Professional Documents
Culture Documents
Tesis Palacios Herrero Pablo
Tesis Palacios Herrero Pablo
ANALYTIC APPROXIMATIONS OF
INTEGRAL TRANSFORMS IN TERMS OF
ELEMENTARY FUNCTIONS:
APPLICATION TO SPECIAL FUNCTIONS
Presented by
Pablo Palacios Herrero
Supervised by
Dr. José L. López Garcı́a
Dr. Pedro J. Pagola Martı́nez
Iruña-Pamplona, January, 2023
Departamento de Estadı́stica,
Informática y Matemáticas,
UNIVERSIDAD PÚBLICA DE NAVARRA
“You may learn much more from a game you lose than from a game you win”
To my parents,
my brother,
my friends,
and Sara.
Acknowledgments
First and foremost, I would like to express my sincere and deepest gratitude to my
supervisors, Professor José Luis López Garcı́a and Professor Pedro Jesús Pagola Martı́nez,
for their unevaluable continuous support of my Ph.D. study and related research. For
their time, patience, motivation, and immense knowledge. Their guidance helped me
since I started in the research field developing my Master Thesis until the final stage of
this dissertation.
Besides my advisors, I would like to acknowledge my colleagues from Universidad
Pública de Navarra, specially Alberto, Anas, Blanca and Joseja for their advices and
the talks during coffee breaks that helped me to break monotomy and keep focus on my
research.
Furthermore, yet importantly, I owe my deepest gratitude to my father, Mr. Antonio
Palacios, and my mother, Mrs. Antonia Herrero. Their constant love and support keep
me motivated and confident. All my successes are thanks to them. My deepest thanks
to my brother, Javier, who always supported me in this adventure. I am forever thankful
for the unconditional love and support from every one of them throughout the entire
thesis process and every day.
I would also like to thank all my friends from Zaragoza. They always believed in me
and helped me to become the person I am today. I am specially thankful to Luis from
my chess club, Palacio de Pioneros, who caught me the research bug.
Similarly, I would like to thank the friends I met in Pamplona, specially my flatmates,
and Sara: For her love and support, specially during the long days of writing this thesis.
She gave me light when everything turned dark and kept me grounded. This thesis could
not have been written without her.
On the other hand, I would like to thank all who helped me during my visit in
Copenhagen, specially to professor Henrik Laurberg Pedersen. He helped me to expand
my horizons and acquire new skills. It was an experience that I will never forget.
Finally, the Universidad Pública de Navarra is acknowledged for its Ph.D. grant that
made this research possible. Likewise, this work was supported by the Ministerio de
Economı́a y Competitividad (MTM2014-52859-P and MTM2017-83490-P) and Universi-
dad Pública de Navarra PRO-UPNA (6158).
From the bottom of my heart... THANKS!!!
v
Contents
Acknowledgments v
Abstract xiii
1 Chapter 1. Introduction 1
1.1 Historical background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 State of the art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Structure of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 Chapter 2. Preliminaries 13
2.1 Multi-point Taylor expansions . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Asymptotic expansions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.1 Watson’s lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2.2 The classical Laplace’s method . . . . . . . . . . . . . . . . . . . 20
2.2.3 A modification of the method of Laplace . . . . . . . . . . . . . . 23
vii
viii Contents
8 Chapter 8. A systematic “saddle point near an end point” asymptotic method 125
8.1 Preparatory results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
8.2 The systematic method “saddle point near an end point” . . . . . . . . . 132
8.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
8.3.1 The confluent hypergeometric function U (c, d, x) for large d and x 139
8.3.2 The confluent hypergeometric function M (c, d, x) for large d and x 140
Appendices 151
D The coefficients of the new expansions for the p+1 Fp function 161
Bibliography 165
List of Figures
3.1 Cases 1-4 concerning the position of the integration interval (0, 1) with
respect to Ω. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2 Case 4: The lemniscate Γr4 is divided into two mirror half-lemniscates. . 34
8.1 Portrait of the minimum of the phase function f (t; α) as a function of α. 129
ix
x List of Figures
5.1 Relative error provided by formula (5.27) for fixed c and several values of t. 81
5.2 Relative error provided by formula (5.27) for c = log(λ0 n/t) and several
values of t. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.3 Relative error provided by formula (5.27) for fixed c and several values of β. 82
5.4 Relative error provided by formula (5.27) for c = log(λ0 n/t) and several
values of β. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.5 Relative error provided by formula (5.27) for fixed c and several values of α. 82
5.6 Relative error provided by formula (5.27) for c = log(λ0 n/t) and several
values of α. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
xi
xii List of Tables
This thesis focuses on the study of new analytical methods for the approximation of
integral transforms and, in particular, of special functions that admit an integral repre-
sentation. The importance of these functions lies in the fact that they are solutions to a
great variety of functional equations that model specific physical phenomena. Moreover,
they play an important role in pure and applied mathematics, as well as in other branches
of science such as chemistry, statistics or economics. Usually, the integrals defining these
special functions depend on various parameters that have a specific physical meaning.
For this reason, it is important to have analytical techniques that allow their computation
in a quick and easy manner. The most commonly used analytical methods are based on
series expansions of local validity: Taylor series and asymptotic (divergent) expansions
that are, respectively, valid for small or large values of the physically relevant variable.
However, neither of them is, in general, simultaneously valid for large and small values
of the variable.
In this thesis we seek new methods for the computation of analytic expansions of
integral transforms satisfying the following three properties:
(a) The expansions are uniformly valid in a large region of the complex plane. Ideally,
these regions should be unbounded and contain the point 0 in their interior.
(b) The expansions are convergent. Therefore, it is not necessary to obtain error bounds
or to study the optimal term to truncate the expansion: the more terms considered,
the smaller the error committed.
We develop a theory of uniform expansions that shows the necessary and sufficient
conditions to obtain expansions of integral transforms fulfilling the three conditions (a),
(b) and (c) above. This theory is applied to obtain new series approximations satisfying
(a), (b) and (c) of a large number of special functions. The new expansions are compared
with other known representations that we may find in the literature to show their ad-
vantages and drawbacks. In contrast to the Taylor and asymptotic expansions, the main
benefit of the uniform expansions is that they are valid in a large region of the complex
plane. For this reason, they may be used to replace the function they approximate (which
is often difficult to work with) when it appears in certain calculations, such as a factor
of an integral or in a differential equation. Since these developments are also given in
terms of elementary functions, such calculations may be carried out easily.
Next, we consider a particularly important case: when the kernel of the integral
transform is given by an exponential. We develop a new Laplace’s method for integrals
xiii
xiv Abstract
that produces asymptotic and convergent expansions, in contrast to the classical Laplace
method which produces divergent developments. The expansions obtained with this new
method satisfy (a) and (b) but not (c), since the asymptotic sequence is given in terms of
incomplete beta functions. Finally, we develop a new uniform asymptotic method ”saddle
point near an end point” which does not satisfy (b) and (c) but, unlike the classical ”saddle
point near an end point” method, allows us to calculate the coefficients of the expansion
by means of a simple and systematic formula.
Chapter 1
Introduction
There is not a formal definition of what a special function is, although they can be defined
as those functions that have a more or less established names and notation and that are
solution to a large variety of functional equations of great significance in mathematics
and physics. They constitute a vast researching field in the area of applied mathematics
as they appear in a large variety of disciplines, not only in mathematics and physics, but
also in others fields such as economy and statistics. The easiest examples of special func-
tions are elementary functions such as the exponential, the logarithm or the trigonometric
functions. However, when we talk about special functions we are typically thinking in
non-elementary functions like, for example, Euler’s gamma function Γ(z), the error func-
tion or Bessel functions. They usually admit and integral or a series representation and,
due to their appearance in a lot of different scientific areas, their analytic approximation
is very important.
The integral representation is indeed valid for non-negative real values of n. Nowadays,
this
R ∞ function is called Euler’s gamma function, it is usually written in the form Γ(z) =
−t z−1
0
e t dt and plays a crucial role in mathematics. G. Michon [87] wrote “Arguably,
the most common special function, or the least special of them. The other transcendental
functions [...] are called special because you could conceivably avoid some of them by
staying away from many specialized mathematical topics. On the other hand, tha gamma
1
2 Chapter 1. Introduction
function Γ(z) is most difficult to avoid”. Just to name some applications, it is used to
evaluate a lot of integrals and compute products, and it is extensively used in quantum
physics, astrophysics and statistics.
On the other hand, Euler also studied the sum of the inverses of integer numbers to
the power of a real number. In particular, he solved the famous Basel problem linking the
summation of the reciprocals of the squares of all the natural numbers with the square of
π. Riemann extended the study to the complex plane by defining the famous Riemann
zeta function [130]. He proved that it is a meromorphic function and found the functional
equation too. He stablished a relation between this function and the distribution of prime
numbers and he also conjectured that all non-trivial zeros of the zeta function have real
part equal to 1/2. This problem, known as Riemann’s conjecture, is still an open problem
considered by many mathematicians to be the most important unsolved problem in pure
mathematics, being the only problem listed in both, Hilbert’s twenty-three Problems
of 1900 and the Millenium Prize Problems created in 2000 by the Clay Mathematics
Institute. Apart from its importance in pure mathematics, the Riemann zeta function has
also many applications in physics. For example, it determines the critical gas temperature
and density for the Bose-Einstein condensation phase transition in a dilute gas [64] and
it is usually used in quantum field theory to perform regularizations to evaluate formally
divergent series.
R
Legendre studied elliptic integrals [63], that is, integrals of the form R(x, y)dx where
R(x, y) is a rational function of the two variables x and y, and y 2 is a polynomial of the
third or fourth degree in x. He was able to show that, in general, those integrals can not
be expressed in terms of elementary functions but he also revealed that, using suitable
transformations, all elliptic integrals may be written in terms of three standard integrals
that are nowadays named after him. The name elliptic integrals is due to the fact that
these integrals are used to evaluate the lenght of an ellipse. However, they can also be
used to obtain the lenght of other plane curves such as an hyperbola or a Bernoulli’s
lemniscate [28]. But, moreover, these special functions play an important role in other
branches of mathematics and physics. For example, they appear in a natural way in
geometrical and statistical problems [51, 128]; they appear in certain electromagnetic
problems [151] and their zeros may be used to obtain an upper bound for the number
of limit cycles of certain hamiltonian systems [146]. Many more applications of elliptic
integrals may be found in [61].
αβ
For his part, Gauss studied an infinite series starting by 1+ 1·γ x+ α(α+1)β(β+1)
1·2·γ(γ+1)
x x+. . .,
that was presented to the mathematical comunity in 1813 [50]. It was the first full system-
atic treatment of such a series and, although some transformation formulas were found by
Gauss, it was clear that, the nowadays called Gauss hypergeometric function, could not
be expressed by means of a closed, simple formula, in terms of elementary funcions. This
function was later studied by Riemann who characterised the second-order differential
equation it satisfies by its three regular singularities (on the Riemann sphere) [129]. A
lot of mathematical functions are special or limiting cases of the Gauss hypergeometric
function. For example, trigonometric functions and the logarithm function, the Kummer
and Bessel functions, the complete elliptic integrals, and a lot of orthogonal polynomials
such as Jacobi, Gegenbauer, Legendre, and Meixner polynomials, are all especial cases
of the Gauss hypergeometric 2 F1 function [104].
Hence, the first investigations on special functions were made in the begining of the
eighteenth century. The number of special functions, their importance and applications
1.2 State of the art 3
started to increase during the next centennial, but it was only on the twentieth century,
with the development of quantum mechanics, that special functions became a vast topic
of interest, as they appear as fundamental solutions of differential equations derived from
spectral problems in atomic potentials. For example, Hermite polynomials are associated
with the eigenfunctions of the harmonic oscillator [57, §18.39] and the spherical harmon-
ics are eigenfunctions of the Laplacian operator in spherical coordinates. Furthermore,
Bessel and hypergeometric functions arise when solving fundamental equations such as
Schrödinger, Helmholtz or Dirac equations by the method of separation of variables [98].
Since then, special functions had appeared not only in pure mathematical problems or in
quantum mechanics, but also in a lot of differents field of study related with engineering
problems, numerical simulations, economic analysis, statistics, optics, chemistry... For
this reason, the list of special function is endless: error functions, Fresnel integrals, Airy
functions, Bessel functions, Kelvin’s function, Struve functions, parabolic cylinder func-
tions, confluent hypergeometric functions, generalized hypergeometric functions, Meijer
G-function, orthogonal polynomials, polylogarithms, theta functions, Mathieu functions,
Lamé functions, catastrophes integrals,... and many more [40, 106, 108]
Due to the significance of special functions, their analytical and numerical approxi-
mation is very important. In this regard, many techniques have been developed: numer-
ical methods for differential equations or integrals, continued fractions, lower and upper
bounds, power series, asymptotics expansions...
On a modern computer, numerical evaluation of special functions or integral trans-
forms can be easily carried out. However, there are a lot of scientific problems where it
is important to have an analytic expression of integral transforms and where a numeri-
cal evaluation would be useless because there are several physical parameters involved;
because we need to know its behavior for small or large values of a certain parameter or
because we have to perform further analytic computations with them. For example and
without the aim of being exhaustive, we give some such situations although we may find
many more in the literature:
The use of numerical methods to approximate integrals for accurate real-time ren-
dering of surfaces lit by non-occluded area light sources leads to results with noise,
hardly compatible with real-time rendering constraints [62].
would be great to have at our disposal expansions valid for small and large values of
the variable. Ideally, those expansions should be valid in unbounded sets of the complex
plane that contain the origin in its interior.
Therefore, we study parametric integrals of the form
Z b
F (z) = g(t)h(t, z)dt, (1.1)
a
for certain functions g(·) and h(·, z). This integral may correspond with an integral
representation of a special function F (z) or it may be a general integral transform. On
the one hand, a lot of special funcions may be cast under this form and even several
special functions are defined via such an integral representation. On the other hand, a
lot of important mathematical integral transforms such as Laplace, Fourier or Stieltjes
transforms are given by the integral (1.1), for a certain kernel h(t, z). Moreover, many
physical quantities in classical mechanics, astronomy or quantum mechanics are also
given by an integral like (1.1).
The most used analytical approximations for integrals of the form (1.1) are Tay-
lor series and asymptotic expansions. The former are found, in the particular case
h(t, z) = h(tz), by considering the Maclaurin expansion of h(tz) at the point t = 0
or by an application of the Frobenius method in the differential equation satisfied by
the special function. In general, the function h(t, z) has some finite singularities and
therefore, the classical Taylor expansion is only valid for small values of the variable |z|.
Contrarily, asymptotic expansions are found by considering an expansion of a factor in
the integrand at its asymptotically relevant point or by an application of Olver’s method
in the differential equation satisfied by the special function. Typically, the integration
interval is unbounded and therefore, when we consider the Taylor expansion of a factor in
the integrand and we interchange summation and integration, we obtain an asymptotic
expansion at the cost of losing the convergence of the expansion.
Consider for example the Bessel function of the first kind Jν (z) [107]. Its power series
expansion is well-known [107, eq. 10.2.2]
ν n−1 k
2 X (−z 2 /4)
Jν (z) = + Rn(0) (ν, z). (1.2)
z k=0
k!Γ(ν + k + 1)
On the other hand, its asymptotic expansion in terms of inverse powers of z is given by
[107, eq. 10.17.3]
r n−1 n−1
πz νπ π X a2k (ν) νπ π X a2k+1 (ν)
Jν (z) = cos z − − −sin z − − +Rn(∞) (ν, z),
2 2 4 k=0 (−z 2 )k 2 4 k=0 z(−z 2 )k
(1.3)
valid for | arg z| ≤ π − δ < π, and a0 (ν) = 1 and, for k = 1, 2, . . .,
0.5 1.0
0.5
0.5
2 4 6 8 10
-0.5 2 4 6 8 10
2 4 6 8 10
Figure 1.1: Approximation of x2 J1 (x) (thicker graphics) given by the Taylor expansion (1.2)
(left), the asymptotic expansion (1.3) (middle) and the uniform expansion (1.4) (right) for
x ∈ [0, 10] and five degress of approximation n = 1, 2, 3, 4, 5 (thinner graphics).
small values of |z|. On the other hand, in [66, §2, Theorem 1] we can find the following
expansion given in terms of elementary functions and that holds uniformly in z in any
fixed horizontal strip of the complex plane:
√
πΓ(ν + 1/2) sin z
J ν (z) = P n−1 (z, ν) − Qn−1 (z, ν) cos z + Rn (z, ν), (1.4)
2(z/2)ν z
where Pn (z, ν) and Qn (z, ν) are rational functions of z and ν [66, eq. 10]. Expansion
(1.4) is valid for large and small values of |z| as the remainder Rn (z, ν) can be bounded
independently of z. The three expansions (1.2), (1.3) and (1.4) are compared in figure
1.1. As we may check from the figure expansion (1.4) is uniformly valid for small and
large values of the variable x and produces expansions that are globally more satisfactory.
As an illustration of the type of expansion (1.4) we may derive, for example, the following
approximation valid for any real x > 0 [66, eq. 7]
4
3x − 140x2 + 360 5(x2 − 18)
15π
J3 (x) = + θ1 (x) x sin x + + θ2 (x) cos x,
2x3 8x6 2x4
where the quantities θ1 and θ2 are uniformly bounded by means of |θ1 (x)| < 0.0062 and
|θ2 (x)| < 0.051.
On the other hand, consider integral transforms of Laplace type, where the kernel
h(t, z) = e−zf (t) is given by an exponential. In the particular case when the phase
function f (t) = t is simply a monomial, we can perform a logarithmic change of variables
to find
Z ∞ Z 1 ∞
−zt z−1
X ck k!
F (z) = e g(t)dt = u g(− log u)du = , (1.5)
0 0 k=0
(z)k+1
where (z)n denotes the Pochhamer’s symbol [4, §5.2(iii)] defined by (z)n = z(z + 1)(z +
2) · · · (z + n − 1) = Γ(z + n)/Γ(z), and ck are certain coefficients related with the Taylor
coefficients of g(− log u) at u = 1 (see Appendix A). The series on the right hand side of
(1.5) is not only convergent in a certain unbounded region of C (at the right of a vertical
line), but it also is an asymptotic expansion of F (z) as z → ∞. Expansions of the form
(1.5) are named factorial series and they were first studied by Newton and Stirling; and
systematically analyzed by Nielsen [96, 97] and Nørlund [99].
Furthermore, many asymptotic methods involve a change of variables to obtain an
asymptotic expansion of an integral. Then, the coefficients of the expansion are defined
1.2 State of the art 7
In this thesis, we face the challenge of designing, if possible, new analytic approxima-
tions of parametric integrals (special functions) satisfying the following three properties:
(a) The expansion is convergent and then, the more terms considered, the better, avoid-
ing the study of the optimal truncation term, inherent to asymptotic (divergent)
series.
(b) They hold uniformly for large and small values of a certain selected variable z, in
contrast to the classical Taylor or asymptotic expansions that are only valid for
small or large values of |z| respectively. That is, the expansion is uniform for z in an
unbounded set of the complex plane that contains the point z = 0. Then, for any
orden of the approximation, its absolute error is bounded independently of z.
(c) The terms of the expansion are elementary functions of z whose coefficients can be
computed by means of an explicit, simple and systematic formula.
CHAPTER 3 of this thesis focuses on the development of a new general theory of uni-
form 1 convergent expansions of integral transforms (1.1) satisfying the three properties
(a), (b) and (c) above.
The main idea is to consider the multi-point Taylor expansion of g(t) at certain se-
lected points in a way that the domain of convergence of the series contains the integration
interval of (1.1), except possibly for one or both end points. When we replace g(t) by
its Taylor series and interchange the order of summation and integration we obtain a
convergent expansion of F (z). Moreover, the independence of z of the function g(t) is
somehow transferred to the remainder of the expansion of F (z) and consequently the
expansion is valid regardless of the value of z. In other words, the expansion is valid for
small and large values of |z|.
We will show that this procedure is not only formal, but rigorous, finding accurate
error bounds that do not depend on the uniform variable z. That is, in section 3.1 we
state the hypotheses that the functions g(·) and h(·, z) must satisfy in order to find a con-
vergent expansion of F (z), given in terms of elementary functions, that holds uniformly
in unbounded subsets of the complex plane that contain the point z = 0. Section 3.2 is
the cornerstone of the new theory of uniform expansions. In it, we perform an accurate
study of the remainder that shows not only the convergence of the new series, but also
its rate of convergence.
Subsequently, a new ingredient for the approximation of integral transforms (1.1) or,
in particular, for integral representations of special functions, is available: their uniform
expansion. These approximations can be of interest to many scientific researchers around
the world that work with special functions and it would be interesting to collect all of them
in a new edition of the famous NIST Handbook of Mathematical Functions [106, 108].
1
The classical meaning of uniform in mathematical analysis is used here. We shall not confuse it with
the notion of uniform used in asymptotics, where the validity of an expansion as a particular parameter
runs along a certain region is studied.
1.3 Structure of the thesis 9
The seed of the idea to derive uniform expansions is [66] where one of the advisors
of this thesis found new expansions of the Bessel functions of the first and second kind,
Jν (z) and Yν (z). Those expansions are convergent and hold uniformly in z in any fixed
horizontal strip of the complex plane. The idea was later applied with sucess by my ad-
visors and collaborators to many more special functions such as the incomplete gamma
and beta functions or the hypergeometric Gauss function [19, 43, 44], obtaining con-
vergent expansions, given in terms of elementary functions, that hold uniformly in z in
unbounded subsets of the complex plane that also contain the point z = 0.
developed in chapter 3. That is, we study integral transforms of the form (1.1) and we
consider the multi-point Taylor expansion not of g(t) as in chapter 3, but of h(t, z), in
selected base points. If we considered the standard Taylor expansion of h(t, z) = h(tz) at
t = 0, we would obtain the well-known power series expansion of the integral transform
(or special function) F (z). The region of convergence of this expansion is typically
small. However, when we consider another, cleverly selected base point for the Taylor
expansion, or even better, a multi-point Taylor expansion of the function h(t, z) = h(tz),
we may find larger regions of convergence that, in many cases, are unbounded and contain
small values of the variable |z|. To illustrate this idea, in this chapter we consider
the generalized hypergeometric p+1 Fp (a, ~b, ~c; z) function and, using the one-, two- and
three-point Taylor expansion of a factor in an integral representation, that we derive in
Appendix C, we obtain different analytical representations of these functions in large
(sometimes unbounded) regions of the complex plane that contain the indented closed
unit disk D∗ = {z ∈ C : |z| ≤ 1, z 6= 1}. The speed of convergence of these expansions is
exponential and they are shown to perform numerically better than other expansions of
the p+1 Fp (a, ~b, ~c; z) function that we may find in the literature. Furthermore, the three
derived expansions are given in terms of rational functions and then, they satisfy the
three properties (a), (b) and (c) described above.
In CHAPTER 7 we consider a specially important case of the integral in the right hand
side of (1.1): when the integration interval is (0, ∞) and the kernel h(t, z) is a exponential
function of the form h(t, z) = e−zf (t) . That is, we study parametric integrals of the form
Z ∞
F (z) = e−zf (t) g(t)dt. (1.6)
0
In the particular case when f (t) = t we can derive a factorial series of F (z) by introducing
a logarithmic change of variables (see (1.5)), which is a convergent series that is also
asymptotic for large z [145, §17.3]. However, generalizations of this results are not given
in the literature. Therefore, we consider integral transforms of the form (1.6) with a
general phase functions f (t) and not simply a monomial of the first degree, and we
derive an expansion of (1.6) that is both, asymptotic for large z and convergent. In
contrast, the classical Laplace’s method introduced in section 2.2.2 and usually used to
approximate integrals of the form (1.6) produces divergent approximations. Moreover,
the coefficients of the new convergent and asymptotic Laplace’s method can be computed
by means of a simple, closed formula, in terms of generalized Bernoulli polynomials. The
derivation of that formula is relegated to appendix A. The price to pay for deriving a
convergent Laplace’s method is that the approximation is given in terms of (sometimes
incomplete) beta functions.
The key point to derive this expansion is to split the phase function f (t) into its
asymptotically dominant part (as z → ∞) and a remainder term, in the form summa-
rized in section 2.2.3. In this way, the study is reduced to the case f (t) = tm , with m ∈ N
(section 7.2) that, after a change of variables, is reduced to compact Mellin transforms,
analyzed in section 7.1. However, as we will see, the conditions on the functions f (t) and
g(t) to apply the results of section 7.2 may be too restrictive. Therefore, in section 7.3
some tricks are performed to enlarge the applicability of the convergent and asymptotic
Laplace’s method. Finally, the method is summarized in the form of a theorem in section
7.4, where it is used to derive a convergent and asymptotic expansion of the hypergeo-
metric confluent U function. In summary, we obtain asymptotic expansions that satisfy
1.3 Structure of the thesis 11
properties (a), (b) and partly (c) above, as the coefficients can be computed by means of
an explicit formula but the asymptotic sequence is no longer given in terms of elementary
functions, but incomplete beta functions.
In CHAPTER 8 we continue the study of integrals (1.1) in the case when h(t, z) =
−zf (t)
e but from another point of view initiated by the advisors of this thesis: the sys-
tematization of classical asymptotic methods of integrals [68, 74, 75].
Using the modified Laplace’s method introduced in [74] and summarized in sec-
tion 2.2.3 simple and systematic formulas for the coefficients of the standard method
of Laplace are found. The same idea was later applied in [68] to obtain a systematic
formula for the coefficients of the “saddle point near a pole” uniform asymptotic method,
whereas in [75] that idea was useful to simplify the computation of the paths of steep-
est descent in the saddle point method. We continue this line of research and revisit
the uniform asymptotic method “saddle point near an end point” [154, Ch. 7, §3], [145,
Ch. 22] where, depending on a certain parameter, the asymptotically dominant point of
the integrand changes of nature from being an interior point to being an end point of
the interval of integration. As a result, an application of the classical Laplace’s method
produces three expansions that are formally different. The classical uniform asymptotic
method “saddle point near an end point” solves this problem by introducing the parabolic
cylinder function into the game. This function encodes the abrupt transition from one
case (the absolute minimum of f (t) is an interior point of (a, b)) to another (the abso-
lute minimum of f (t) occurs at one end point, say t = a). However, as in the classical
Laplace’s method, simple and explicit formulas for the coefficients of the expansions are
not known. Therefore, following the ideas of the modified Laplace’s method of section
2.2.3 we obtain an asymptotic expansion of the integral (1.6) uniformly valid as that pa-
rameter varies. As in the classical method, the expansion is given in terms of parabolic
cylinder functions but, on the other hand, the coefficients are given by means of an
explicit, closed formula that is systematically computable.
Finally, in CHAPTER 9 we draw some conclusions of the results of this thesis and we
suggest some future work to continue this investigation.
Thoughout the thesis, for any complex variable z, arg z ∈ (−π, π] denotes its main
argument and logarithms and fractional powers are assumed to take their principal value.
The symbol z → ∞ or the statement“large z”mean |z| → ∞ with fixed arg z. The symbol
bac denotes the greatest integer less than or equal to a whereas the symbol dae denotes
the least integer greater than or equal to a. All the computations of the thesis have been
performed with the symbolic program Wolfram Mathematica 12, version 12.1.1.0.
Chapter 2
Preliminaries
Taylor and multi-point Taylor expansions of an analytic funcion f (z) are valid in a
bounded domain unless the function f (z) is entire. In general they are not useful when
the variable z becomes extremely large. However, in a lot of physical problems there are
quantities that take values far from the rest. For example, in kinematics the speed of
light is extremely large compared with other speeds; in certain central potential problems
the central mass is extremely heavier than the other masses; the viscosity coefficient
of a slightly viscous fluid is small compared with other constants in fluid mechanics...
Moreover, small coefficients in boundary value problems lead to singularly perturbed
problems [37, 38, 110] which appear, for example, in fluid or gas dynamics [84, 147], heat
transfer [5, 6], atmospheric pollution [89]... Then, it is important to have solutions to
this kind of problems. And these solutions should be easy to compute, in terms of the
asymptotic parameter, as that parameter approaches a certain selected value (possibly
infinity).
13
14 Chapter 2. Preliminaries
Most of the physical or engineering problems where asymptotic parameters occur are
described by differential or integral equations whose solution admits an integral represen-
tation. For this reason, the theory of asymptotic approximations is divided in two main
areas: the asymptotic approximation of solutions of differential equations [105, 149] and
the asymptotic approximation of integrals [14, 145, 154].
The former can be directly applied to a large number of problems but, as there is
not an explicit representation for the solution to be approximated, it is usually harder
to apply. This is so because, apart from the differential equation, we need some more
extra information to consider the right solution to be estimated. On the other hand, the
solution to a lot of real problems can be written in terms of an integral by using integral
transforms (Laplace, Fourier, Mellin, etc.) or Green’s function techniques. Even further,
the problem itself may be given by means of such an integral representation.
In this thesis, we focus on the approximation of functions given by means of an integral
transform. For this reason, the asymptotic theory of integrals will be used throughout
the next chapters. The main aim of the asymptotic theory of integrals is to obtain an
approximation of a function F (z) given in the form of a parametric integral, valid as a
certain selected (asymptotic) parameter approaches a limit value. Typically, the analytic
expression of F (z) is difficult to work with and then, we seek expansions given in terms
of simpler (ideally elementary) functions.
In the later chapters of this thesis, multi-point Taylor expansions and asymptotic
expansions will be used. Then, for clarity in the exposition, the main results of these
topics, adapted to the applications that we need in the next chapters, are given in the
next two sections.
where pk (z) are polynomials of degree m − 1 that can be represented by the following
Lagrange-type formula:
m Qm
s=1,s6=j (z − zs )
X
pk (z) := ak,j Qm , (2.2)
j=1 s=1,s6 = j (zj − zs )
being ak,j coefficients related with the first k derivatives of f (z) evaluated at the base
points:
" # m
" #
1 dk f (z) X 1 dk−1 f (z)/(z − zj )
ak,j := Qm + Qm .
k! dz k s=1,s6=j (z − zs )
k
l=1,l6=j
(k − 1)! dz k−1
s=1,s6 = l (z − zs)
k
z=zj z=zl
(2.3)
2.1 Multi-point Taylor expansions 15
2 2 2
z1
1 1 1
0 0 0
z1 z2 z1 z2
z2
-1 -1 -1
-2 -2 -2
-2 -1 0 1 2 -2 -1 0 1 2 -2 -1 0 1 2
Figure 2.1: Different shapes of the Cassini oval of convergence Om for m = 2 depending on
the relative size of the “radius” r and the position of the two base points z1 and z2 .
1.5 1.5 1.5
0.0 0.0
z2 0.0 z1 z2
z1 z2 z3 z1
Figure 2.2: Different shapes of the lemniscate domain Om for m = 3 depending on the relative
size of the “radius” r and the position of the base points z1 , z2 and z3 .
The multi-point Taylor remainder rn (z) in (2.1) is given by the following Cauchy’s
integral formula
Qm n I
s=1 (z − zs ) f (w)dw
rn (z) = Qm n
, (2.4)
2πi C (w − z) s=1 (w − zs )
where C is a simple closed loop contained in Ω that encircles the points z1 , z2 , . . . , zm and
z in the counterclockwise direction.
Expansion (2.1) is uniformly and absolutely convergent inside a lemniscate Om (see
figures 2.1 and 2.2), where
( m
) (m )
Y Y
Om ≡ z ∈ Ω : |z − zs | < r , r ≡ inf |w − zs | .
w∈C\Ω
s=1 s=1
If m = 1, the domain O1 is a disk and we obtain the classical Taylor series expansion; if
m = 2 the domain O2 is a Cassini oval (see figure 2.1); for m greater than 2 the lemniscate
takes different shapes depending on the relative size of the “radius” r and the position of
the base points (see figure 2.2).
In the next chapters we will need to approximate a function f (z) by means of multi-
point Taylor expansions in a way that the lemniscate of convergence Dr contains the
16 Chapter 2. Preliminaries
Im(w)
Ω
Im(w)
D Ω Im(w)
x
Re(w)
x D D
0 1/2 1 Re(w)
x
Re(w)
0 1 0 1/2 1
x
x
Ω
Figure 2.3: Lemniscates of convergence Dr with base points t1 , . . . , tm with 0 ≤ t1 < t2 <
. . . tm ≤ 1 for different number of base points m = 1 (left), m = 2 (middle) and m = 3 (right).
In the tree pictures, the crossed points represent the singularities of the given functions f (w)
and define the value of the maximal “radius” ρ. Left: f (w) = (2w + 1)−1 . Disk of radius
r0 = 1/2 < r < ρ = 1 centered at 1/2. Middle: f (w) = (5 − 16w + 16w2 )−1 . Cassini oval of
radius r0 = 1/4 < r √ < ρ = 5/16 and√foci at 0, 2 −1
√1. Right: f (w) = (20w − 8w + 1) Lemniscate
of radius r0 = 1/(12 3) < r < ρ = 13/(20 10) and foci at 0, 1/2, 1.
In every example, the lemniscate Dr contains the interval (0, 1) but avoids the singularities
of f (w). The more base points t1 , t2 ,..., tm the lemniscate Dr has, the better it avoids the
singularities of f (w), as Dr becomes more and more thinner (always containing the interval
(0, 1)) as we can see in the sequence of examples (left)-(middle)-(right).
interval [0, 1] while being contained in the analyticity region Ω of the function f (z). In
order to obtain it, we choose m different real points 0 ≤ t1 < t2 < . . . < tm ≤ 1 and we
define the quantities (m )
Y
r0 := sup |t − ts | (2.5)
t∈(0,1) s=1
and ( )
m
Y
ρ := inf |w − ts | . (2.6)
w∈C\Ω
s=1
the coefficients Ak,j can be computed using the following recurrent algorithm:
Define
m
Y
U (z) = (z − ts )
s=1
Then, the coefficients Ak,j , j = 0, 1, 2, . . . , m − 1 are the solution to the following Van-
dermonde linear system
1 t1 t21 . . . tm−1
1 Ak,0 φk (t1 )
1 t 2 t22 . . . tm−1
2
Ak,1 φk (t2 )
.. .. = .. , (2.9)
.. .. .. ..
. . . . . . .
1 tm t2m m−1
. . . tm Ak,m−1 φk (tm )
The matrix in the system (2.9) is a Vandermonde matrix which is, in general, ill-
condicionated and some numerical care should be taken when using this method. How-
ever, in many practical examples, we take 1, 2 or at most 3 base points for the multi-point
Taylor expansion that are, in general, well-distribuited along the interval [0, 1], say t1 = 0,
t2 = 1/2 and t3 = 1. Therefore, numerical problems should not occur and this recurrent
algorithm to compute the coefficients Ak,j may be useful.
On the other hand, the function f (z) usually satisfies a certain differential equation of
the first or second order. Thence, it is usually possible to obtain a recurrence relation to
compute the coefficients Ak,j by replacing f (z) and its derivatives by their corresponding
multi-point Taylor expansion and equating the coefficients of equal powers.
Two particular values of L should be highlighted. On the one hand, if L = 0 we say that
f is of order less than g, and we write f (z) = o (g(z)), as z → z0 . That is,
On the other hand, if L = 1 we say that f and g are asymptotically equal at z0 and we
denote it by f (z) ∼ g(z). In other words,
For properties on the O-symbols, we refer to [105, Ch. 1] or [119, Ch. 1].
These three symbols are also used in the definition of an asymptotic expansion of
a function. First, if {φn (z)}∞
n=0 is a sequence of continuous functions such that, for all
n = 0, 1, 2, . . .
φn+1 (z) = o (φn (z)) , as z → z0 , (2.10)
we say that {φn (z)}∞ n=0 is an asymptotic sequence or an asymptotic scale. Then, if
∞
{φn (z)}n=0 is such an asymptotic sequence as z → z0 and f (z) is a function that, for
every fixed n = 0, 1, 2, . . ., verifies
n−1
X
f (z) − ak φk (z) = O (φn (z)) , as z → z0 , with ak ∈ C, ∀k, (2.11)
k=0
As it has been said in the introduction of this chapter, the theory of asymptotic
expansions is divided in two areas: the approximation of solutions to differential equations
and the approximation of integrals. In this thesis, we are only interested in the latter
case as we are interested in the approximation of integral transforms or special functions
2.2 Asymptotic expansions 19
being λ a real or complex number with <λ > 0 and µ > 0 a positive number.
(iii) The integral (2.13) converges for sufficiently large positive values of <z.
Then the integral F (z) admits the following asymptotic expansion, for large |z|
X∞ Γ n+λ
µ
F (z) ∼ gn n+λ , z→∞ (2.14)
n=0 z µ
π
in the sector | arg z| ≤ 2
− δ < π2 .
The sector of validity of expansion (2.14) can be extended when more information on
the analytic properties of the function f (t) is available [145, pp. 14]. Generalizations of
the lemma of Watson may be found in [154, Ch. 1, §5]. Moreover, in [145, Ch. 2] we can
find a Watson’s lemma for loop integrals.
20 Chapter 2. Preliminaries
Example 1. Consider the function F (z) = ez E1 (z), where E1 (z) denotes the exponential
integral [142, eq. 6.2.2] Z ∞ −zt
e
F (z) = dt.
0 t+1
We have the expansion
∞
1 X
g(t) = = (−1)n tn , |t| < 1,
t + 1 n=0
that is convergent in any disk Dr (0) centered at t = 0 of radius r < 1. Thus, in this
particular example, expansion (2.14) reads [142, eq. 6.12.1]
∞
X (−1)n n!
F (z) ∼ , z → ∞. (2.15)
n=0
z n+1
We highlight that the series (2.15) is asymptotic, for large z, but it is not convergent
for any value of z ∈ C. The divergence of expansion (2.15) may be regarded as a penalty
for interchanging summation and integration when the disk of convergence of the Taylor
series expansion of g(t) does not contain the integration interval1 . In any case, this
is not important in asymptotics and only the two properties (2.10) and (2.11) for the
asymptotic sequence and the remainder matters.
where either a or b or both may be infinite. For the sake of simplicity, we assume that
the asymptotic variable x > 0 is positive and large, and also that the functions f (t) and
g(t) are real for real t. On the other hand, we allow a branch point at the end integration
point t = a given by the factor (t − a)s−1 . We also assume that f (t) and g(t) are smooth
enough for t ∈ (a, b).
Laplace [59] made the observation that, for large x, the peak value of e−xf (t) occurs at
the point t = t0 where f (t) attains its minimum value. For large x, the peak is very sharp
and the main contribution of the integrand to the integral comes from the neighborhood
of that point. Then, Laplace derived the approximation
s
2π
F (x) ∼ g(t0 )(t0 − a)s−1 e−xf (t0 ) , as x → ∞,
xf 00 (t0 )
which is the first term of a full asymptotic expansion that was rigorously found by Erdélyi
[39, Ch. 2, §4].
1
Even if g(t) was an entire function, the convergence of the asymptotic expansion found after an
application of Watson’s lemma is not assured.
2.2 Asymptotic expansions 21
Erdélyi’s approach is the following: If the function f (t) has a finite number of minima
and maxima, break up the integral in a finite number of integrals and reverse the sign
of x if necessary so that in each integral f (t) reaches its minimum value at one (and
only one) point t = t0 ∈ [a, b). Then, we can perform a change of variables and apply
Watson’s lemma to the resulting integral to find an asymptotic expansion of the integral
F (x) in (2.16).
From now on, the analysis strongly depends on the position of the absolute minimum
t0 of f (t) with respect to the integration interval [a, b). We distinguish three different
cases: (i) the absolute minimum of f (t) is an interior point t0 ∈ (a, b) and it is a simple
zero of f 0 (t); (ii) t0 = a and it is a simple zero of f 0 (t); (iii) the function f (t) is strictly
increasing in [a, b) and therefore the absolute minimum of f (t) in [a, b) occurs at t0 = a,
with f 0 (a) > 0.
(i) The absolute minimum of f (t) occurs at the interior point t0 ∈ (a, b), with f 0 (t0 ) = 0
and f 00 (t0 ) > 0. Then, the graph of f (t) near the point t = t0 is that of a parabola.
For this reason, we introduce in (2.16) the change of variable t 7→ u given by
f (t)−f (t0 ) = u2 and we extend the integration limits to ±∞, since the contribution
of the tails up to ±∞ to the integral are exponentially small. We find
Z +∞
−xf (t0 ) 2
F (x) = e e−xu h(u)du, (2.17)
−∞
the integral occurs near the point u = 0 and, under some mild hypotheses on the
functions f (t) and g(t), it can be shown that the function h(u) has an asymptotic
expansion at u = 0 of the form
∞
X
h(u) ∼ hn un , as u → 0+ ,
n=0
(ii) The absolute minimum of the phase function f (t) occurs at one of the end integra-
tion points, say t = a, and it is a simple, critical point of f (t). We can repeat step
by step the computations of the previous case, but with two differences:
1. After the change of variable f (t) − f (t0 ) = u2 and adding the exponentially
negligible tails, the lower integration point in the integral (2.17) is not u = −∞
but u = 0.
22 Chapter 2. Preliminaries
of the form ∞
X
h(u) ∼ hn un+s−1 , as u → 0+ .
n=0
(iii) The function f (t) is strictly incresing in [a, b) and therefore it attains its absolute
minimum at the end point t = a, but with f 0 (a) > 0. Now, the shape of the graph
of f (t) near the absolute minimum t = a is no longer a parabola, but a straight line.
Thus, we no longer perform a quadratic transformation as in the previous cases, but
a change of variable of the form f (t) − f (t0 ) = u. Then, except for exponentially
small terms, we have that
Z ∞
−xf (t0 )
F (x) = e e−xu h̄(u)du, (2.20)
0
Then, Watson’s lemma yields the asymptotic expansion [154, Ch. 7, §3]
∞
−xf (t0 )
X Γ (n + s)
F (x) ∼ e hn , as x → ∞. (2.21)
n=0
xn+s
This exposition of the Laplace’s method is more intuitive than rigorous. For details
on the assumptions for the functions f (t) and g(t), the sector of validity of the expansions
for complex values of z and also for a proof of these results, we refer to [105, Ch. 3, §7],
[145, Ch. 3], [154, Ch. 2, §1]. If the integral (2.16) was defined along a complex path,
the saddle point method should instead by applied [105, Ch. 4, §§6 and 7], [145, Ch. 4],
[154, Ch. 2, §4].
The method of Laplace can not be direcly applied to this integral, as the absolute ex-
tremum of the integrand is attained at a point that depends on the asymptotic variable
2.2 Asymptotic expansions 23
z. However, after a straightforward change of variables, the method can be applied and
the following expansion may be derived [4, eq. 5.11.3]
1/2
−z z 2π 1 1
Γ(z) ∼ e z 1+ + + ... , as z → ∞,
z 12z 288z 2
whose first term is the famous Stirling’s formula [134] for the factorial of a large integer
number. A simple and explicit formula for the coefficients between brackets of the right
hand side of this asymptotic expansion may be found in [91].
Note that the expansions obtained by the classical Laplace’s method are, in general,
divergent. Even if the function h(u) admits not only an asymptotic series at u = 0,
but a Taylor series, the radius of convergence is usually small and does not contain
the integration interval (see example 1). In other words, the price to pay for deriving
expansions (2.18), (2.19) or (2.21) by interchanging summation and integration is that
the expansions usually diverge. Consequently, accurate error bounds and a study of the
optimal truncation term are usually needed in practice, although many times the first
term of the expansion gives enough information on the asymptotic behavior of F (x).
Also, if the asymptotic variable x is large enough, we can obtain as many precision as
required.
In any of the three cases (i), (ii) or (iii), the asymptotic sequence is easy to compute
as it is nothing but inverse powers of the asymptotic variable x, that is, it is a Poincaré-
type asymptotic sequence. In contrast, the coefficients hn of the expansion involve the
reversion of the series of h(u) or h̄(u), a function defined implicitly by the change of
variables f (t) − f (t0 ) = u2 or f (t) − f (t0 ) = u. Traditional textbooks on asymptotics
[105, 154] do not give explicit formulas for those coefficients, but rather an indication on
how to compute the first few. Nevertheless, some more or less explicit representations of
these coefficients can be found in the literature: Perron’s formula [125] gives them in terms
of the derivatives of an explicit function whereas Lauwerier [60] derives a formula for the
coefficients in the form of an integral of an exponential, a power function and a polynomial
pn that must be computed recursively using a n−terms recurrence that involves integrals.
On the other hand, Wojdylo [152, 153] found a recurrence formula for the coefficients of
the Laplace expansion in terms of partial ordinary Bell polynomials. More recently, based
on those expressions, Nemes [92] obtained a formula for the coefficients given in terms of
ordinary potential polynomials. In any case, those formulas to calculate the coefficients hn
after an aplication of Laplace’s method involve the computation of combinatorial objects
whose complexity increases with the number of terms considered. In order to circumvent
this problem, a modification of Laplace’s method for integrals was introduced in [74].
by considering the first non-constant and non-vanishing term of the Taylor expansion of
f (t) at the asymptotically relevant point t = t0 (the absolute minimum of f (t) in [a, b]).
More precisely, the modification of the Laplace’s method introduced in [74] works as
follows: we consider again the integral (2.16) being (a, b) a real (may be unbounded)
interval and f (t), g(t) smooth enough functions. Without loss of generality we assume
that s ∈ (0, 1]. We also assume that f (t) is real-valued and that it has a unique absolute
minimum at t = t0 ∈ [a, b). We also require that the functions f (t) and g(t) have a
Taylor series expansion2 at t = t0 with common radius of convergence r > 0. Let m ∈ N
denote the first non-vanishing derivative of f (t) at t = t0 , that is, f 0 (t0 ) = f 00 (t0 ) =
. . . f (m−1) (t0 ) = 0 and f (m) (t0 ) > 0. If t0 ∈ (a, b) is an interior point, then m is even. We
consider also the remainder function
f (m) (t0 )
fm (t) := f (t) − f (t0 ) − (t − t0 )m
m!
and we split the exponential in the form
f (m) (t0 )
(t−t0 )m
e−xf (t) = e−xf (t0 ) e−x m! e−xfm (t) .
For large x > 0, the main contribution of the integrand to the integral comes from the
neighborhood of the point t = t0 . But moreover, the contribution of the factor e−xfm (t) to
f (m) (t0 )
(t−t0 )m
the integral is subdominant compared with the contribution of e−x m! . There-
fore, we re-write the integral (2.16) in the form
Z b f (m) (t0 )
−xf (t0 ) (t−t0 )m
F (x) = e e−x m! h(t, x)dt, h(t, x) = e−xfm (t) g(t)(t − a)s−1 . (2.22)
a
bn/mc
X xk f (m) (t0 ) k n−mk
X
xf (t0 )
hn (x) = e Aj (x)Bn−mk−j . (2.24)
k=0
k! m! j=0
2
This condition may be relaxed and require only that both, f (t) and g(t) have an asymptotic expan-
sion at t = t0 .
2.2 Asymptotic expansions 25
When we introduce the expansion (2.23) into (2.22), interchange summation and inte-
gration and extend (if necessary) the integration interval up to ±∞, we obtain [74, eq.
19]
∞
" n+s #
n
X 1 + (−1) β n + s m! m
F (x) ∼ e−xf (t0 ) hn (x) Γ , as x → ∞,
n=0
m m f (m) (t0 )x
(2.25)
with β = 0 if t0 = a and β = 1 otherwise. Expansion (2.25) is in fact an asymptotic
expansion of F (x) for large x, although this is not obvious as the coefficients hn (x) depend
on the asymptotic variable x. However, in [74] it has been shown that the coefficients
hn (x) are polynomials in x of degree b np c, where p > m denotes the first non-vanishing
derivative of f (t) at t = t0 after the m−th derivative. Thus, the terms of the expansion
n+s
(2.25) between brackets are O xbn/pc− m , as x → ∞. They do not constitute a genuine
Poincaré sequence that decrease monotonically with the order n of the approximation in
the form x−n , but in the form of a sawtooth.
Chapter 3
Uniform Expansions of
Integral Transforms
As it has been pointed out in the introduction 1, in this chapter we face the challenge of
designing a new theory of uniform approximation of parametric integrals satisfying the
following three properties:
(a) The expansion holds uniformly for large and small values of a certain selected variable
z.
To this end, we consider an integral transform of a function g(t) with kernel h(t, z) of
the form Z b
F (z) = h(t, z)g(t)dt, (3.1)
a
where we assume that (a, b) is a bounded or unbounded interval and h(·, z)g(·) is inte-
grable on (a, b). We also assume that g(t) is analytic in a region Ω ⊂ C that includes the
open set (a, b) ⊂ Ω. It could be that the function F (z) depends on other extra variables
that we assume to be included in h and/or g. So, we omit their explicit indication as we
are interested in the function F (z) as a function of a single (selected) variable z. Then,
most of the special functions of the mathematical physics can be written in the form of
an integral transform like (3.1) (see [106, 108]).
Moreover, after splitting the integration interval if necessary and performing an affine
change of variables on (3.1), we may assume that the integration interval (a, b) is whether
[0, 1] if [a, b] was bounded or [0, ∞) if (a, b) was unbounded. That is, we consider that
the function F (z) in (3.1) is one of the following two:
Z 1 Z ∞
F (z) = h(t, z)g(t)dt, or F (z) = h(u, z)g(u)du, (3.2)
0 0
with (0, 1) ⊂ Ω in the first case and (0, ∞) ⊂ Ω in the second case. But moreover,
after a change of variable that transforms the unbounded interval (0, ∞) into the unit
interval (0, 1) we can reduce the study of the second case of (3.2) to the first one. Such a
27
28 Chapter 3. Uniform expansions of integral transforms
change of variable can be, for example, logarithmic: u = − log t. In this case, we identify
h(− log t, z)/t with h(t, z) and g(− log t) with g(t). Therefore, without loss of generality,
we study integral transforms of a function g(t) with kernel h(t, z) of the form
Z 1
F (z) = h(t, z)g(t)dt, z ∈ D, (3.3)
0
where D is a bounded or unbounded region of the complex plane. In general, the logarith-
mic change of variables u = − log t or any other change that transforms the unbounded
interval (0, ∞) into the interval (0, 1) makes the point t = 0 and/or t = 1 in (3.3) a singu-
lar point of g(t). Then, in the later analysis of the integral (3.3) we consider four different
cases, depending on whether the end points 0 and 1 of the integration interval are regular
or singular points of the function g(t), which, as we will see, will be an essential aspect of
the analysis. In other words, we consider four different situations concerning the position
of the end points 0 and 1 with respect to Ω:
1. [0, 1] ⊂ Ω. That is, none of the end points are singular points of g(t).
4. (0, 1) ⊂ Ω but [0, 1] ( Ω. Both end points 0 and 1 are singular points of g(t).
The main idea to derive a uniform expansion of the integral F (z) in (3.3) is to con-
sider the multi-point Taylor expansion of the function g(t) at selected points such that
(0, 1) ⊂ Dr ⊂ Ω, where Dr is the lemniscate of convergence of the chosen multi-point
Taylor expansion of g(t). Then, replace g(t) in (3.3) by that expansion and interchange
summation and integration. Under some hypotheses that we state in the next section
and with accurate bounds for the remainder, we will show that this method produces an
expansion of F (z) satisfying the following three properties:
(a) The expansion is uniform 1 for z in an unbounded subset D ⊂ C such that D̄ contains
the point z = 0. Note that we are using the classical meaning of uniform of the
mathematical analysis: for any order n of the approximation Fn (z), the absolute
error |F (z) − Fn (z)| ≤ Mn for any z ∈ D, with Mn independent of z. That is, the
expansion is valid for both large and small values of |z|.
(b) The expansion is convergent. That is, the remainder of the expansion, Rn (z) :=
F (z) − Fn (z), vanishes as n → ∞, for any z ∈ D.
To ensure that properties (a), (b) and (c) are satisfied, the following assumptions on the
functions h(t, z) and g(t) are considered:
(i) The function h(t, z) can be dominated by an integrable function H(t) on [0, 1] in
the sense that |h(t, z)| ≤ H(t) for all z ∈ D and for all t ∈ [0, 1].
1
This use of uniform expansion shall not be confused with the notion of “uniform expansions” used in
asymptotics. In asymptotics, the term “uniform expansion” is used to refer to an expansion of a function
of several variables that remains valid as a certain selected parameter crosses a critical value.
29
(ii) The function g(t) is analytic in a region Ω ⊂ C that contains the open set (0, 1) ⊂ Ω.
R1
(iii) The moments of the function h(t, z), M [h(·, z); k] := 0
h(t, z)tk dt are elementary
functions of z.
Then, when 0 ∈ D the requirement (i) implies the property (a). The requirement (ii)
assures the property (b) and the requierement (iii) guarantees the property (c).
Compare the above mentioned idea with the convergent (Taylor) expansions or the
asymptotic expansions of F (z). Roughly speaking, the former are found by replacing
h(t, z) in the integrand of the integral (3.3) that defines F (z) by its Taylor expansion, as
a function of z, at z = 0; whereas the latter are found in a similar way, but considering
the asymptotic expansion of g(t) at the asymptotically relevant point of h(t, z). Then, in-
terchanging summation and integration we find expansions of F (z) that are, respectively,
convergent or asymptotic. However, in general, those expansions do not hold uniformly
in z. They are useful and accurate for small (convergent) or large (asymptotic) values
of the variable z. But their validity is local, that is, they are useless when we want to
evaluate the function away from small (convergent) or large (asymptotic) values of z.
On the other, an expansion satisfying the three properties (a), (b) and (c) listed above
is globally more satisfactory as it is valid in a large region of the complex plane, that
contains small and large values of the selected variable z (see figure 1.1 in chapter 1 for
an illustration of this discussion).
Moreover, imagine that a certain special function F (z) admitting an integral repre-
sentation of the form (3.3) appears in a certain computation like for example a factor of
the integrand of an integral or in a differential equation. In general, neither the Taylor se-
ries expansions nor the asymptotic approximations can replace the function F (z) in that
computation, as they do not hold in a large enough region (range of integration, domain
of the differential equation...). In contrast, an expansion satisfying the three properties
(a), (b) and (c) is uniformly valid in a large set of the complex plane and therefore it can
be used to replace the function F (z) in that computation. As the approximation is given
in terms of elementary functions, it should be easier to work with it instead of dealing
with the function F (z), whose analytical expresion can be more involved.
Using the above described procedure, uniform expansions of some particular special
functions F (z) (for particular functions h(t, z) and g(t) in the right hand side of (3.3))
and satisfying the properties (a), (b) and (c) above) have been obtained in [66] (Bessel
functions), [19] (incomplete gamma function), [20] (confluent hypergeometric function),
[44] (incomplete beta function), [43] (Gauss hypergeometric function), [69] (generalized
hypergeometric functions p−1 Fp and p Fp ) and [21, 22] (symmetric elliptic functions). A
brief summary of these expansions is relegated to the next chapter but, as we will see,
in all these examples, the derived expansions satisfy properties (a), (b) and (c) listed
above. In this chapter of the thesis, we show that this is not a coincidence but those are
just particular examples of a general theory: the theory of uniform expansions of integral
transforms.
This general theory has been developed in the paper [77].
30 Chapter 3. Uniform expansions of integral transforms
3.1 Hypotheses
In this section, we clearly set the hypotheses required for the two factors h(t, z) and g(t) in
the integrand of (3.3) pointed out above. As we have already mentioned, we assume that
g(w) is an analytic function in an open region Ω that contains the integration interval
[0, 1] except, possibly, for an integrable singularity at w = 0 and/or at w = 1. More
precisely:
Hypothesis 1. We assume that g(w) is analytic in an open region Ω that contains
the interval (0, 1) and the function f (w) := w1−σ (1 − w)1−γ g(w), with 0 < σ, γ ≤ 1 is
bounded in Ω. In particular, according to the four different situations that we may find
concerning the position of the end points 0 and 1 with respect to Ω, we have:
σ =1 and γ = 1, in case 1, [0, 1] ⊂ Ω.
σ <1 and γ = 1, in case 2, (0, 1] ⊂ Ω but [0, 1] ( Ω.
σ =1 and γ < 1, in case 3, [0, 1) ⊂ Ω but [0, 1] ( Ω.
(0, 1) ⊂ Ω but [0, 1] ( Ω.
σ <1 and γ < 1, in case 4,
We have also mentioned above that the function h(t, z) is majorized by an integrable
function H(t) on [0, 1], for all z ∈ D. In particular:
Hypothesis 2. We assume that |h(t, z)| ≤ Htα (1 − t)β for (t, z) ∈ [0, 1] × D, with
H > 0 a positive number independent of z and t and the parameters α and β satisfying
α + σ > 0 and β + γ > 0.
Observe that it is natural to assume this form for the bound of the function h(t, z) as
the product h(·, z)g(·) must be integrable in [0, 1]. Also, we could consider H not to be a
constant, but an integrable function on [0, 1] independent of z. The price to pay by doing
so is too high, as working with that H(t) function would be much more cumbersome
and the advantage would not be that great, as the requirement “H constant” is usually
enough in practice.
Hypothesis 3. We can choose m different base points 0 ≤ t1 < t2 < . . . < tm ≤ 1
such that the lemniscate (2.7) of convergence of the multi-point Taylor expansion of g(t)
satisfies (0, 1) ⊂ Dr ⊂ Ω, that is, r0 ≤ r ≤ ρ (see equations (2.5) and (2.6) in chapter 2
for the definition of the minimal and maximal “radius” r0 and ρ respectively).
Note that, as it is explained in [80] (in a different context), if (0, 1) ⊂ Ω we can always
find an appropriate selection of the base points t1 , . . . , tm to assure that r0 ≤ ρ and we
can take a “radius” r such that r0 ≤ r ≤ ρ or, equivalently, (0, 1) ⊂ Dr ⊂ Ω.
Recall the expresion (2.1) for the multi-point Taylor expansion of an analytic function
g(z):
n−1
"m #k
X Y
g(z) = pk (z) (z − ts ) + gn (z), (3.4)
k=0 s=1
where pk (z) is a polynomial in z of degree m − 1 (see section 2.1 in chapter 2) and gn (z)
is the multi-point Taylor remainder that can be represented by the following Cauchy’s
integral (see (2.4))
Qm n I
s=1 (z − ts ) g(w)dw
gn (z) := Qm n
, z ∈ Dr . (3.5)
2πi Γ (w − z) s=1 (w − ts )
3.2 Analysis of the remainder gn (t) 31
When w = t ∈ (0, 1) and [0, 1] ⊂ Ω (case 1) we may choose the base points ts ,
s = 1, 2, . . . , m such that Γ ⊂ Ω with ε = 0. In the other three cases (cases 2, 3 and
4), when [0, 1] ( Ω it is not possible to take ε = 0 as the contour Γ would contain the
singular points t = 0 and/or t = 1. Notwithstanding, for any t ∈ (0, 1) and small enough
ε, the contour Γ encircles not only the base points t1 , t2 , . . . , tm but also the point t. But,
moreover, in any of the four cases, and in particular in cases 2, 3 and 4, the integral in the
right hand side of (3.5) is a (constant) function of ε as it is the integral of an integrable
function. Therefore, we can consider the limit ε → 0 and consider that the “radius” of the
lemniscate in the above integral equals r. Then, this limit lemniscate Dr ( Ω although
Dr \ {0, 1} ⊂ Ω. Then, we may consider that the “radius” r of the lemniscate boundary
Γ that defines the integration contour in (3.5) is such that the interval [0, 1] ⊂ Dr .
Case 2: We Q have to choose Q t1 > 0. The “radius” r2 of the lemniscate Dr2 must
satisfy r2 := m t
s=1 s ≥ r 0 > m
s=1 (1 − ts ). This condition assures that the interval
(0, 1] ⊂ Dr2 . The point 0 ∈ / Dr2 but 0 ∈ D̄r2 .
Case 3: We Q have to choose tm <Q 1. The “radius” r3 of the lemniscate Dr3 must
satisfy r3 := s=1 (1 − ts ) ≥ r0 > m
m
s=1 ts . This condition assures that the interval
[0, 1) ⊂ Dr3 . The point 1 ∈/ Dr3 but 1 ∈ D̄r3 .
Case 4: We have toQchoose 0Q < t1 < tm < 1. The “radius” r4 of the lemniscate Dr4
must satisfy r4 := s=1 ts = m
m
s=1 (1 − ts ). This condition assures that the interval
(0, 1) ⊂ Dr4 . The points 0, 1 ∈
/ Dr4 but 0, 1 ∈ D̄r4 .
With this election of the base points t1 , . . . , tm and of the “radius” rj of the lemniscate
Drj in each of the four cases j = 1, 2, 3, 4, and taking into account the above observation
32 Chapter 3. Uniform expansions of integral transforms
Case 1 Case 2
Case 3 Case 4
Figure 3.1: A possible election of appropriate base points in the case m = 3. We have taken:
Case 1 (top left) t1 = 0, t2 = 1/2, t3 = 1 and r1 = 1/20 > r0 . Case 2 (top right) t1 = 1/10,
t2 = 1/2, t3 = 1 and r2 = t1 t2 t3 = 1/20 = r0 . Case 3 (bottom left) t1 = 0, t2 = 1/2, t3 = 9/10
and r3 = (1 − t1 )(1 − t2 )(1 − t3 ) = 1/20 = r0 . Case 4 (bottom right) t1 = 1/14, t2 = 1/2,
t3 = 13/14 and r4 = t1 t2 t3 = 13/392 = r0 .
on the integration contour Γ, we assume that the integration path is, in each of the four
cases, the boundary of the corresponding lemniscate Drj , that is, Γrj := ∂Drj . In figure
3.1 we show a possible selection of the base points and “radius” with m = 3 in the four
different cases 1-4.
In the remaining of this section we find sharp bounds for the remainder gn (t) in all
four cases 1-4.
t ∈ [0, t1 ]. We also use hypothesis 1, namely, that f (w) := w1−σ g(w) is bounded in
Dr2 , with 0 < σ < 1. Then, from (3.5) we find, for t ∈ [0, t1 ]:
Qm
|t − ts |n |wσ−1 f (w)| (t1 − t)n |wσ−1 f (w)|
I I
1 s=1
|gn (t)| ≤ Qm dw ≤ M dw,
2π s=1 |ts |
n
Γr2 |w − t| tn1 Γr2 |w − t|
where the integration contour is the scaled lemniscate boundary Γr2 /t:
t1 t2 tm r2
Γr2 /t = w ∈ C : w − w− ··· w − = m .
t t t t
For any t > 0, the most left point of this scaled lemniscate is the point w = 0
and the most right point is the point w = t0 /t, where t0 is the most right point of
the lemniscate boundary Γr2 . In the limit t → 0 the scaled lemniscate boundary
Γr2 /t becomes the imaginary axis traversed downwards. In this path, we have that
|f (wt)| ≤ M0 , with M0 > 0 independent of w ∈ Γr2 /t and t > 0 and the integral
|wσ−1 |
I
dw, 0<σ<1 (3.7)
Γr2 /t |w − 1|
Case 3. It is similar to the previous studied case 2, but interchanging the roles
of the points w = 0 and w = 1 and the lemniscates Dr2 and Dr3 . That is, case 3
becomes case 2 after the change of variables w 7→ 1−w, considering the factorization
g(w) = (1 − w)γ−1 f (w) instead of g(w) = wσ−1 f (w), the “radius” r3 instead of r2
and reversing the order of the points t1 , t2 , . . . , tm . Then, for t ∈ [0, tm ] we obtain
a bound similar to the bound obtained in the case 1:
M
|gn (t)| ≤ , t ∈ [0, tm ],
an
with M > 0 and a > 1 independent of t and n. On the other hand, for t ∈ [tm , 1)
we find
(t − tm )n (1 − t)γ−1
|gn (t)| ≤ M , t ∈ [tm , 1), (3.9)
(1 − tm )n
with 0 < γ < 1 given in hypothesis 1 and M > 0 independent of t and n.
34 Chapter 3. Uniform expansions of integral transforms
C1
Re(w)=1/2
t
0 t1 t3 1
C0
Figure 3.2: Case 4: The lemniscate Γr4 is divided into two mirror half-lemniscates C0 and C1
after cutting Γr4 with the vertical line <w = 1/2. In this example, m = 3, t1 > 0, t2 = 1/2 and
t3 = 1 − t1 .
with 0 < σ < 1 given in hypothesis 1 and a certain constant M > 0 independent
of t and n.
A similar analysis shows that, for t ∈ [tm , 1),
(t − tm )n (1 − t)γ−1
|gn (t)| ≤ M , t ∈ [tm , 1),
(1 − tm )n
These sharp bounds for the remainder gn (t) in the four analyzed cases 1-4 are going
to be used in the next section, where we derive the main result of this chapter: a uniform
convergent expansion of the integral F (z) given in (3.3).
Theorem 3.3.1. Assume that hypotheses 1-3 of section 3.2 for the functions g(t) and
h(t, z) and for the base points t1 , t2 , . . . , tm of the multi-point Taylor expansion of g(t)
hold. Then, for n = 1, 2, 3, . . . , and z ∈ D,
Z 1 n−1 m−1
X X
F (z) = h(t, z)g(t)dt = Ak,j M [h(·, z); k, j] + Rn (z), (3.12)
0 k=0 j=0
where Ak,j are the multi-point Taylor coefficients of the function g(t) at the base points
t1 , t2 , . . . , tm (see formulas (2.1), (2.2), (2.8) and (2.9) in chapter 2), and M [h(·, z); k, j]
are the multi-point moments:
" m
#k
Z 1 Y
M [h(·, z); k, j] := h(t, z) (t − ts ) tj dt (3.13)
0 s=1
On the other hand, the remainder Rn (z) may be bounded in the form
1 n! Γ(α + σ) n! Γ(β + γ)
|Rn (z)| ≤ M H n + A +B , (3.14)
a Γ(n + α + σ + 1) Γ(n + β + γ + 1)
36 Chapter 3. Uniform expansions of integral transforms
with
(0, 0) in case 1,
(1, 0) in case 2,
(A, B) :=
(0, 1) in case 3,
(1, 1) in case 4.
Furthermore, the constants H > 0 and M > 0 are independent of t and n and have been
respectively introduced in hypothesis 2 of section 3.1 and in the analysis of the remainder
gn (t) in section 3.2. The parameters α, β, σ and γ are defined in hypotheses 1 and 2
and a > 1, derived in section 3.2, is independent of t and n.
Therefore, expansion (3.12) is uniformly convergent for z ∈ D in any of the four
cases 1-4, but we note that the rate of convergence strongly depends on the case under
consideration: the convergence is of exponential type in case 1, that is, Rn (z) = O(a−n );
whereas it is of power type in the remaining cases, Rn (z) = O(n−δ ), for some δ > 0.
More precisely, as n → ∞,
Proof. Consider the multi-point Taylor expansion (2.1) of the function g(t) at the base
points t1 < t2 < . . . < tm , with representation (2.8) of pk (t), that converges in the
lemniscate Dr , with (0, 1) ⊂ Dr ⊂ Ω and r = r1 , r2 , r3 or r4 depending on the case 1-4
under consideration. Replace g(t) on the left hand side of (3.12) by that multi-point
Taylor expansion and interchange summation and integration to find the right hand side
of (3.12), with
Z 1
Rn (t) := h(t, z)gn (t)dt. (3.16)
0
Note that, by hypotesis 2, the multi-point moments (3.13) of h(t, z) exists. It remains to
show that the remainder Rn (z) can be bounded as in (3.14). The analysis is different in
the four cases 1-4.
Case 1. From the analysis of the remainder gn (t) in section 3.2, case 1, it is clear
that |gn (t)| ≤ M a−n for all t ∈ [0, 1], with M > 0 and a > 1 independent of t and
n. Then, if we introduce this bound into (3.16) we get (3.14) with A = B = 0.
Case 2. From the analysis of the remainder gn (t) in section 3.2, case 2, we have
that |gn (t)| ≤ M a−n , with M > 0 and a > 1 independent of t and n, for t ∈ [t1 , 1]
n σ−1
and we have that |gn (t)| ≤ M (t1 −t)tn t , for t ∈ (0, t1 ]. Then, we write
1
Z t1 Z 1
Rn (z) = h(t, z)gn (t)dt + h(t, z)gn (t)dt.
0 t1
We introduce the above bounds of gn (t) and the bound h(t, z) ≤ Htα given in
hypothesis 2 to find
M H t1
Z Z 1
n α+σ−1 MH α+σ MH
|Rn (z)| ≤ n (t1 −t) t dt+ n = M Ht1 tα+σ−1 (1−t)n dt+ n .
t1 0 a 0 a
Using the definition of Euler’s beta function [4, eq. 5.12.1] and the fact that 0 <
tα+σ
1 ≤ 1 we find the bound (3.14) with A = 1, B = 0.
3.3 A uniform convergent expansion of the integral F (z) 37
In the first integral, we have |gn (t)| ≤ M a−n , with M > 0 and a > 1 independent
)n (1−t)γ−1
of t and n, whereas in the second integral we have |gn (t)| ≤ M (t−tm(1−t m)
n . Then,
it follows that
Z 1
MH MH
|Rn (z)| ≤ n + (1 − t)β+γ−1 (t − tm )n dt
a (1 − tm )n tm
Z 1−tm
MH MH
≤ n + tβ+γ−1 (1 − tm − t)n dt
a (1 − tm )n 0
Z 1
MH β+γ
≤ n + M H(1 − tm ) tβ+γ−1 (1 − t)n dt.
a 0
The derivation of bound (3.14) follows the same steps that in the previous cases: In
n σ−1
the first integral we use the bound |gn (t)| ≤ M (t1 −t)tn t ; in the second one we have
1
n γ−1
|gn (t)| ≤ M a−n whereas in the last integral |gn (t)| ≤ M (t−tm(1−t
) (1−t)
m)
n holds. Then,
using hypothesis 2 and these bounds, formula (3.14) with A = B = 1 follows.
Finally, using the asymptotic behavior of the quotient of two gamma functions [4, eq.
5.11.12] in the right hand side of (3.14) we find the convergence rate (3.15).
Some remarks are to be done:
Remark 3.3.2. The bound (3.14) of the remainder Rn (z) of the expansion given in the
right hand side of (3.12) is independent of z. Then, no matter how small or large |z|
is, the expansion (3.12) is valid as long as z ∈ D. That is, the expansion given in the
theorem is valid uniformly on z, for all z ∈ D. And it is also convergent, as shown by
the rate of convergence (3.15).
Remark 3.3.3. In cases 2-4 the proof is more involved than in case 1. We could have
repeated step by step the simpler proof of case 1, but with a = 1. Then, that proof would
not have shown the convergence of expansion (3.12) as the parameter a in (3.14) would
not have been large enough (> 1). Therefore, the more involved proof in cases 2-4 is
neccesary to show the convergence.
Remark 3.3.4. In case 1 when both end points of the integration interval [0, 1] are con-
tained in the region Ω of analyticity of the function g(w), the convergence of the expan-
sion (3.12) is of exponential type, that is, the remainder Rn (z) = O(a−n ), with a > 1.
In contrast, when one or both end points are not contained in Ω, the scenario is worse:
Expansion (3.12) is still convergent, but the convergence is only of power type, that is,
the remainder Rn (z) = O(n−δ ), for some δ > 0.
38 Chapter 3. Uniform expansions of integral transforms
As shown in (3.15) the value of δ is related with the behavior of the integrand in the
left hand side of (3.12) at the end points t = 0 and/or t = 1. That behavior is somehow
transferred to the remainder Rn (z) in the right hand side of expansion (3.12). This
analysis reminds [46, 47] where it is proved that the asymptotic behavior of a function
g(t), with integrable singularities in the unit disk, is transferred to the asymptotic behavior
of its standard Taylor coefficients.
Remark 3.3.5. It follows from (3.14) that, the larger α and β are, the faster the con-
vergence of the expansion of F (z) in cases 2-4 is. In case 1, we have considered the
possibility α = β = 0 because the bound |h(t, z)| ≤ Htα (1 − t)β with α and/or β > 0 does
not mean any improvement in the convergence speed of expansion (3.12): regardless α
and/or β vanish or not, we would derive formula (3.14) for the remainder in case 1.
The same applies to the second integral in the analysis of case 2: we have used the
bound |h(t, z)| ≤ H because the bound |h(t, z)| ≤ Htα (1 − t)β with α and/or β > 0 does
not mean any improvement in the convergence speed of expansion (3.12). On the other
hand, in the first integral of case 2 we have considered β = 0, but α ≥ 0 as the factor
tα accelerates the speed of convergence but the factor (1 − t)β does not. The situation is
analogous in cases 3 and 4.
Remark 3.3.6. Given an integral, we have to factorize the integrand in two functions g(t)
and h(t, z) satisfying hypotheses 1-3 in section 3.1, being z the uniform variable. For
example, consider the following integral representation of the incomplete gamma function
[116, eq. 8.2.1] Z 1
γ(a, z) = z a e−zt ta−1 dt, <a > 0.
0
It is straightforward to check that the inmediate choice g(t) = ta−1 and h(t, z) = e−zt
satisfies the hypotheses 1-3 in D = {z ∈ C : <z ≥ Λ} for any Λ ∈ R. Indeed, we
can take H = max{e−Λ , 1}, α = β = 0 and γ = 1, σ = a − d<ae + 1 (recall that
0 < σ ≤ 1). Then, we are in case 1 if a ∈ N and in case 2 if a ∈ / N. In the former case
the incomplete gamma function reduces to an elementary function whereas in the latter
the point t = 0 is a branch-point of g(t). Therefore, we assume that a ∈ / N and we apply
directly the theorem 3.3.1 with t1 = 1/2 as the only base point of the Taylor expansion of
g(t) (m = 1) to find the uniform expansion [19, eq. 8]. However, with this election of the
functions g(t) and h(t, z), using (3.15) in theorem 3.3.1 we would find that the remainder
Rn (z) = O(n−(a−d<ae+1) ). This exponent is too small (between −1 and 0) and does not
agree with the order given in [19, eq. 13] for the remainder: We are losing information
on the speed of convergence. The reason is that the parameter σ must satisfy 0 < σ < 1.
Taking into account the previous remark 3.3.5, we note that the greater α and β are,
the faster the convergence is. Then, a clever factorization of the integrand is the following:
take g(t) = ta−d<ae and h(t, z) = e−zt td<ae−1 . Considering D = {z ∈ C : <z ≥ Λ} for any
Λ ∈ R, it is easy to show again that hypotheses 1-3 hold by taking the only base point
t1 = 1/2 (and therefeore m = 1). We again find H = max{e−Λ , 1}, β = 0 and γ = 1 but
α = d<ae − 1 and σ = <a − d<ae + 1, case 2 assuming that a ∈ / N. That is, the value of
σ is the same that on the previous choice, but the value of α is greater. In this way, we
find a uniform and convergent expansion of the incomplete gamma function, similar to
the one given in [19, eq. 8]. The remainder satisfies Rn (z) = O(n−σ−α ) = n−<a , which
agrees with [19, eq. 15] and shows the speed of convergence of the expansion without
losing information, in contrast to the first election of g(t) and h(t, z) taken above in this
remark.
3.3 A uniform convergent expansion of the integral F (z) 39
The restriction 0 < σ < 1 is neccesary from a theoretical point of view as it is needed
to assure that the integral (3.7) in the proof of theorem 3.3.1 is finite. On the other hand,
we are not losing generality as we can take σ to be the fractional part of the exponent a
in ta−1 as in the example of remark 3.3.6. But moreover, in many practical situation, as
in the case of the incomplete gamma function γ(a, z) analyzed in [19], we may not split
the exponent a as the sum of its integer and fractional parts as we did in the second part
of remark 3.3.6 and simply take g(t) = ta−1 and h(t, z) = e−zt . As we have an explicit
expression for the functions g(t) and h(t, z) we may derive and explicit bound for the
remainder [19, eq. 13] that shows Rn (z) = O(n−<a ) instead of the implicit bound (3.14)
that would lead to Rn (z) = O n−(a−d<ae+1) , after applying theorem 3.3.1, where we lose
information on the speed of convergence.
Remark 3.3.7. As we have just seen in the example of the incomplete gamma function,
a bound of the form |h(t, z)| ≤ Htα (1 − t)β with H > 0, constant, and α and/or β ≥ 0 is
common in practice. We could make the theory a bit more general by relaxing hypothesis 2
replacing “H > 0 a positive number independent of z and t” by “H an integrable function
of the variable t”. On the one hand, the price to pay would be a more involved derivation
of the uniform bounds of Rn (z). On the other hand, the theses of theorem 3.3.1 would be
essentially the same ones, with a slight modification of the form of (3.14). Therefore, we
do not consider this possible generalization and we simply assume that h(t, z) is bounded
in the form given in hypothesis 2.
In the next chapter of the thesis, we derive new uniformly convergent expansions of
many special functions of the mathematical physics having an integral representation of
the form (3.12), by means of a direct application of theorem 3.3.1. We also compare
the uniform approximation with the classical power series and asymptotic expansions of
those special functions.
Chapter 4
New Uniform Expansions of
Some Special Functions
This chapter of the thesis is devoted to the application of the theory of uniform approxi-
mation of integral transforms developed in chapter 3 to some particular special functions
of the mathematical physics having an integral representation of the form (3.2). That
is, we are going to obtain new convergent expansions, given in terms of elementary func-
tions, of many special functions. Those expansions hold uniformly in large sets of the
complex plane that may be unbounded and contain both, large and small values of the
uniform variable. Several such representations have already been found by the advisors
of this thesis and collaborators. For example, in [66] we can find an expansion of the
Bessel function of the first kind Jν (z) that holds uniformly in z in any fixed horizontal
strip of the complex plane |=z| ≤ Λ whereas in [20] we can find two uniform expansions
of the confluent hypergeometric M (a, b, z) function uniformly valid for either <z ≥ 0 or
<z ≤ 0. On the other hand, in [43] we can find an expansion of the Gauss hypergeometric
function 2 F1 (a, b, c; z) that holds uniformly for z in an unbounded region of the complex
plane that depends on a certain angle θ ∈ (0, π/2] that, in the limit θ → 0 becomes the
cutted plane C \ [1, ∞).
In all the examples that we are going to analyze, we follow the notation of chapter
3. In particular, we identify the functions g(t) and h(t, z) to classify the problem in
one of the cases 1-4 analyzed in theorem 3.3.1. Specially important is the value of the
parameters α, β, σ and γ in order to estimate the speed of convergence of the expansion.
However, a remark on the value of those parameters must be done: In chapter 3 the
parameters σ and γ were restricted to the unit interval, 0 < σ, γ ≤ 1. This is not a
restriction on the applicability of theorem 3.3.1 as in the event of a branch point in the
integrand, say ta , we could simply take σ to be the fractional part of a. However, by doing
so and using theorem 3.3.1 we may be losing information on the speed of convergence
of the approximation. In order not to lose that information, we can consider that the
analytic function ta−σ is part of the function h(t, z) (see remark 3.3.6). Both choices are
possible and produce similar, but slightly different approximations.
On the other hand, in the particular examples of special functions, we have an explicit
expression for the functions g(t) and h(t, z) and many times it is possible to find an
explicit expression for the remainder of the Taylor expansion of g(t) and, with it, to
derive an explicit bound for the remainder Rn (z) of the uniform approximation of the
41
42 Chapter 4. New uniform expansions of some special functions
special function. Typically, that bound would show, as n → ∞, that Rn (z) = O (n−a )
and not only Rn (z) = O (n−σ ), being σ the fractional part of a, as a direct application
of theorem 3.3.1 would indicate. The same applies for γ if the function g(t) had an
integrable singularity at t = 1. Then, in many examples of special functions below, such
an explicit bound for the remainder has been found in the literature.
In summary, we consider several special functions and we obtain new convergent ex-
pansions that are valid in a large region of the complex plane of a selected variable. Some
of these expansions have been derived by the advisors of this thesis and by collaborators
of my researching group, but I did not contribute to its derivation. Therefore, I have de-
cided to cite the approximations while omitting many details in their derivation. On the
other hand, in the derivation of the uniform approximations of the special functions in
which I took part (elliptic functions, Struve function, incomplete gamma Γ(a, z) function)
more details are given.
To illustrate the kind of expansions that we can derive using theorem 3.3.1, we give the
first terms of the uniform approximation, graphics and numerical tables for the different
special functions that we analyze. In this way, we can see that the derived expansions are
given in terms of elementary functions, that they are convergent and also their uniform
features. We also compare the uniform approximations with the well-known power series
and asymptotic expansions of the corresponding special function.
valid for all z ∈ C. We want to derive a convergent expansion of Hν (z) uniformly valid for
z ∈ D, with D a large set of the complex plane to be determined. To this end, we identify
g(t) = (1 − t2 )ν−1/2 and h(t, z) = sin(zt). Obviously, |h(t, z)| ≤ cosh(=z) ≤ cosh(Λ) for
all t ∈ (0, 1) and z in any fixed horizontal strip |=z| ≤ Λ, with Λ > 0. Therefore, we find
α = β = 0. On the other hand, for general values of ν ∈ C, the function g(t) has two
singular points, located at t = ±1.
We consider the Taylor series expansion of g(t) at an arbitrary point λ ∈ [0, 1/2).
This choice of λ assures that the integration interval (0, 1) is contained in the disk of
convergence of the Taylor series of g(t). Therefore, we can apply theorem 3.3.1 (case 3)
to the integral in the right hand side of (4.1) to obtain the expansion
√ ν n−1
πΓ(ν + 1/2) 2 X (1/2 − ν)k
Hν (z) = (1 − λ)ν−k−1/2 Fk (λ, z) + Rn (λ, z), (4.2)
2 z k=0
k!
R1
where Rn (λ, z) := 0 sin(zt)gn (t, λ)dt being gn (t, λ) the Taylor remainder of g(t) at t = λ
after n terms of the approximation.
4.1 The Struve function Hν (z) 43
On the one hand, the functions Fk (λ, z) are elementary functions of z. They are given
by
k 1 k
d2j
1 − cos z
Z
X k k−j 2j k
X k k−j
Fk (λ, z) = (−λ) t sin(zt)dt = (−1) λ
j=0
j 0 j=0
j z dz 2j
j j−1 2l+1
k
" #
2l l l
X k (2j)! X z (−1) X z (−1)
= (−1)k λk−j 2j+1 1 − cos z − sin z ,
j=0
j z l=0
(2l)! l=0
(2l + 1)!
(4.3)
where, for z = 0, the right hand side of this formula must be understood in the limit
sense.
On the other hand, for any λ ∈ [0, 1/2) and according to (3.15) in theorem 3.3.1
(case 3), the remainder Rn (λ, z) satisfies, as n → ∞, Rn (λ, z) = O (n−γ ), being γ the
fractional part of ν + 1/2. However, as we will see below, we are losing information on
the speed of convergence and the remainder actually satisfies Rn (λ, z) = O n−(ν+1/2) ,
As gn (t, λ) is the Taylor remainder of g(t) and t ∈ (0, 1) is contained in the disk of
convergence of the Taylor series, we have that
∞
X (1/2 − ν)k
gn (t, λ) = (1 − λ)ν−k−1/2 (t2 − λ)k
k=n
k!
n + 1/2 − ν, 1 t2 − λ
(1/2 − ν)n ν−n−1/2 2 n
= (1 − λ) (t − λ) 2 F1 .
n! n+1 1−λ
√
Introducing the last formula into (4.4), splitting that integral at t = λ and after some
computations, we may find the bound
| (1/2 − ν)n |
|Rn (λ, z)| ≤ cosh(Λ) [G1 (ν, λ, n) + G2 (ν, λ, n)] (4.5)
n!
with
√ n
√
π n! λ <ν+1/2 1, 1 + <ν
G1 (ν, λ, n) = λ(1 − λ) 2 F1 λ ,
2 Γ(n + 3/2) 1−λ n + 3/2
(1 − λ)<ν+1/2
1, 1/2, 1/2 + <ν
G2 (ν, λ, n) = 3 F2 1−λ .
1 + 2<ν n + 1, <ν + 3/2
n+1/2
−1/2 λ
For λ ∈ [0, 1/2] we have that, as n → ∞, G1 (ν, λ, n) = O n 1−λ
and
G2 (ν, λ, n) = O (1). Therefore, from (4.5) and using the asymptotic behavior of the
−(ν+1/2)
quotient of two gamma functions, we find Rn (λ, z) = O n , as n → ∞, for any
λ ∈ [0, 1/2].
44 Chapter 4. New uniform expansions of some special functions
0.6
0.0005
0.5 0.4
0.0004
0.4 0.2
0.0003
0.3
2 4 6 8 10
0.0002
0.2
-0.2
0.1 0.0001
-0.4
0 2 4 6 8 10 0 2 4 6 8 10 12 14
Figure 4.1: Real (left) and imaginary (middle) parts of the approximations of the function on
the left hand side of (4.2) for ν = 11/5 (black, dashed) provided by the Taylor expansion [117,
eq. 11.2.1] (green), the asymptotic expansion [117, eqs. 11.2.5, 11.6.1 and 10.17.4] (blue) and
the uniform expansion (4.2) for λ = 0 (red), for z ∈ [0, 10eiπ/10 ] after n = 4 terms. The figure
on the right shows their relative errors for n = 10, ν = 18/5 and real z ∈ [0, 15]. The behavior
is similar for other values of ν and z.
0.008
5
10
0.006
2 4 6 8 10
0.004
5
-5
0.002
-10
2 4 6 8 10
0 2 4 6 8 10
Figure 4.2: Real (left) and imaginary (middle) parts of the approximations of the Bessel
function z2 J1 (z) (black, dashed) provided by the Taylor expansion [107, eq. 10.2.2] (green),
the asymptotic expansion [107, eq. 10.17.3] (blue) and the uniform expansion [66, Theorem 1]
(red), for z ∈ [0, 10eiπ/4 ] after n = 4 terms. The figure on the right shows their absolute errors
for n = 10, ν = 1 and real z ∈ [0, 10]. The behavior is similar for other values of ν and z.
kind
1
2(z/2)ν
Z
Jν (z) = √ (1 − t2 )ν−1/2 cos(zt)dt, <ν > −1/2,
πΓ(ν + 1/2) 0
valid for all z ∈ C. Following the notation of chapter 3, we identify g(t) = (1 − t2 )ν−1/2
and h(t, z) = cos(zt) and we consider any fixed horizontal strip of the complex plane
D := {z ∈ C : |=z| ≤ Λ}, for any Λ > 0. Then, for all z ∈ D and t ∈ (0, 1) we have
|h(t, z)| ≤ sinh(Λ). Therefore α = β = 0. On the other hand, the function g(t) is the
same as in the Struve Hν (z) function. Hence, we are in case 3 of the four cases analyzed
in chapter 3 and we have σ = 0 and γ = ν + 1/2, with (A, B) = (0, 1). We can apply
theorem 3.3.1 to find a convergent expansion of the Bessel function of the first kind that
is uniformly convergent in z ∈ D. This expansion is given in [66, eq. 9]. The remainder
of the expansion satisfies Rn (z, ν) = O n−ν−1/2 , as n → ∞ [66, eq. 12] (see (3.15) in
theorem 3.3.1). As an illustration, we obtain the following approximation valid for x > 0
[66, eq. 7]:
4
3x − 140x2 + 360 5(x2 − 18)
15π
J3 (x) = + θ1 (x) x sin x + + θ2 (x) cos x,
2x3 8x6 2x4
with |θ1 (x)| ≤ 0.0062 and |θ2 (x)| ≤ 0.051.
In figure 4.2 and table 4.2 the convergent and uniform features of the approximation
are shown. The uniform approximation is also compared with the well-known power
series expansion [107, eq. 10.2.2] and asymptotic expansion [107, eq. 10.17.3] of the
Bessel function.
Remark 4.2.1. For ν = m + 1/2 with m = 0, 1, 2, . . . the Bessel function is an elemen-
tary function and the uniform expansion [66, Theorem 1] after n terms is exact for n
sufficiently large (n ≥ m + 1) as the bound [66, eq. 12] for the remainder vanishes.
46 Chapter 4. New uniform expansions of some special functions
0.6 0.6
0.008
0.4 0.4
2 4 6 8 10 2 4 6 8 10 0.004
-0.2 -0.2
0.002
-0.4 -0.4
-0.6 -0.6
0 2 4 6 8 10
Figure 4.3: Real part (left) and imaginary part (middle) of the approximations of the incom-
plete gamma function z −a γ(a, z) (black, dashed) provided by the power series expansion [116,
eq. 8.7.1] (green), the asymptotic expansion [116, eqs. 8.2.3 and 8.11.2] (blue) and the uniform
expansion [19, Theorem 1] (red), for z ∈ [0, 10i] after n = 4 terms, for a = 3/2. The figure on
the right shows their relative errors for n = 10, a = 9/2 and real z ∈ [0, 10]. The behavior is
similar for other values of z and a.
Instead, for convenience in the analysis of chapter 5 we are interested in the derivation
of a convergent expansion of Γ(a, x) with x > 0 uniformly valid in a region of the complex
a−plane that contains the interval (−∞, 0).
Therefore, we assume that x > 0 and we consider the integral definition of the in-
complete gamma function [116, eq. 8.2.2]. We perform a change of variables of the form
s 7→ u given by s = xeu to obtain
Z ∞ Z ∞
a−1 −s u
Γ(a, x) := s e ds = x a
eau e−xe du.
x 0
We have found the unbounded integration interval [0, ∞). For this reason, and following
the steps of chapter 3, we perform a further change of variables u 7→ t given by u = − log t.
We get
Z 1
Γ(a, x) = x a
t−a−1 e−x/t dt. (4.7)
0
−x/t
As we are considering a as the uniform parameter, we identify g(t) = e t and h(t, a) =
t−a . It is easy to check that the hypotheses of theorem 3.3.1 are satisfied and to derive a
convergent expansion of Γ(a, x) valid uniformly in <a < 0 by directly applying theorem
3.3.1. Again, for the convenience of the analysis in chapter 5 we derive the expansion
with an bound for the remainder.
The function g(t) admits the following representation, in terms of the Laguerre poly-
(0)
nomials Ln (x) ≡ Ln (x) [57],
n−1
e−x/t −x
X
g(t) = =e Lk (x)(1 − t)k + rn (x, t), (4.8)
t k=0
48 Chapter 4. New uniform expansions of some special functions
which directly follows from the generating function of Ln (x) [57, eq. 18.12.13], namely,
w ∞
ex w−1 X
= Ln (x)wn , |w| < 1,
1 − w n=0
by setting w = 1 − t.
Thus, expansion (4.8) is convergent for |1 − t| < 1. As a consequence, the remainder
rn (x, t) satisfies
∞
X
−x
rn (x, t) = e Lk (x)(1 − t)k .
k=n
On the other hand, introducing (4.8) in (4.7) and using the integral representation of the
beta function we obtain
n−1
X
Γ(a, x) = e−x xa Lk (x)B(1 − a, k + 1) + Rn (a, x), (4.9)
k=0
with
Z 1 Z 1 ∞
X
a −a a −x −a
Rn (a, x) := x t rn (x, t)dt = x e t Lk (x)(1 − t)k dt.
0 0 k=n
In [57, eq. 18.14.8] we find the inequality e−x/2 |Ln (x)| ≤ 1, valid for 0 ≤ x < +∞. Then
Z 1 ∞
X Z 1
<a −x/2 −<a <a −x/2
|Rn (a, x)| ≤ x e t k
(1 − t) dt = x e t−<a−1 (1 − t)n dt.
0 k=n 0
Using one more time the integral representation of the beta function, we find the
bound
|Rn (a, x)| ≤ e−x/2 x<a B (−<a, n + 1) ≤ e−x/2 x<a Γ (−<a) n<a . (4.10)
The second inequality above follows from [4, eq. 5.6.8]. Indeed, we can write
Γ(−<a)Γ(n + 1) Γ(−<a)nΓ(n)
B (−<a, n + 1) = = ≤ Γ(−<a)nn<a−1 = Γ(−<a)n<a .
Γ(n + 1 − <a) Γ(n + 1 − <a)
The results of this section for the incomplete gamma Γ(a, z) function will be used in
chapter 5. For this reason, we summarize them in the form of a corollary:
Corollary 4.4.1. Let a ∈ C and x ∈ R such that x > 0 > <a. Then, for any n ∈ N,
the incomplete gamma function Γ(a, x) admits the representation (4.9) whose remainder
Rn (a, x) can be bounded in the form (4.10). In particular, the expansion is convergent
for any complex number a in the semi-plane <a ≤ Λ < 0, for any fixed Λ < 0, and the
convergence rate is of power order. That is, Rn (a, x) = O (na ), as n → ∞.
Some plots comparing the Taylor series expansion at a = 0 of the incomplete gamma
funcion Γ(a, x), the asymptotic expansion for large a [116, eqs. 8.11.4 and 8.2.3] and the
uniform approximation (4.9) are shown in figure 4.4. On the other hand, in table 4.4 the
uniform and convergent features of the expansion (4.9) are exhibited.
4.5 The hypergeometric confluent M function 49
0.12 0.10
-10 -8 -6 -4 -2
0.10
0.08
-0.005
0.08
0.06
0.06
-0.010
0.04
0.04
-0.015 0.02
0.02
Figure 4.4: Real part (left) and imaginary part (middle) of the approximations of the incom-
plete gamma function x−a Γ(a, x) (black, dashed) provided by the Taylor series expansion at
a = 0 (green), the asymptotic expansion [116, eqs. 8.11.4 and 8.2.3] (blue) and the uniform
expansion (4.9) (red), for a ∈ [−10eiπ/6 , 0] after n = 4 terms, for x = 3/2. The figure on the
right shows their relative errors for n = 4, x = 3 and a ∈ [−10eiπ/4 , 0]. The behavior is similar
for other values of x and a.
and we assume that neither a nor b − a is an integer, as in those cases the M (a, b, z)
function can be expressed as incomplete gamma functions and the analysis is reduced to
section 4.3.
We want to obtain a convergent expansion for M (a, b, z) uniformly valid in the variable
z, with z in a large region of C. Following chapter 3 we have two main choices for the
functions g(t) and h(t, z).
We can take h(t, z) = ezt ta−1 and g(t) = (1 − t)b−a−1 . Then, we are in case 3 of
theorem 3.3.1 and we may find α = a − 1, β = 0, σ = 1 and γ = b − a, with
(A, B) = (0, 1). Considering t = 0 as the only base point of the Taylor series
of g(t) and using theorem 3.3.1 we obtain expansion [20, eq. 10] that was first
found by K.E. Muller [90]. The remainder term of the expansion is of the order
Rn (z) = O (n−a ), as n → ∞. However, the moments of h(t, z) are not given in
terms of elementary functions, but rather in terms of incomplete gamma functions.
50 Chapter 4. New uniform expansions of some special functions
0.10
0.1
1.0
0.08
0.8 -10 -8 -6 -4 -2
-0.1
0.06
0.6
-0.2
0.4 0.04
-0.3
0.2 0.02
-0.4
-10 -8 -6 -4 -2
-0.5 -10 -8 -6 -4 -2 0
Figure 4.5: Real part (left) and imaginary part (middle) of the approximations of the hyper-
geometric M (a, b, z) function (black, dashed) provided by the power series expansion [103, eq.
13.2.2] (green), the asymptotic expansion [103, eq. 13.7.2] (blue) and the uniform approxima-
tion [20, eq. 21] (red) for z ∈ [−10eiπ/4 , 0], a = 2.1 and b = 4.2 when we truncate the series
after n = 4 terms. The figure in the right shows their relative error after n = 10 terms for
a = 1.2, b = 3.3 and real z ∈ [−10, 0].
with |ε(z)| ≤ 0.223 in <z ≤ 0. For z = 0 the right hand side of this formula must
be understood in the limit sense. The bound for the remainder ε(z) follows from
[20, eq. 28].
In figure 4.5 and table 4.5 the accuracy of the uniform expansion is exhibited. It
is compared with the well-known power series expansion [103, eq. 13.2.2] and the
asymptotic expansion [103, eq. 13.7.2] of the confluent M function.
Moreover, the previous expression is uniformly convergent for z with <z ≤ Λ, for
any fixed Λ ∈ R. But, using Kummer’s transformation M (a, b, z) = ez M (b −
a, b, −z) [103, eq. 13.2.39] and the previous expansion for the function M , we can
obtain an expansion for M that holds uniformly in z for <z ≥ Λ for any Λ ∈ R [20,
eq. 29].
We take g(t) = (− log t)a−1 (1 − log t)b−a−1 and h(t, z) = tz−1 . For any fixed Λ > 0 we
define the region D := {z ∈ C : <z ≥ Λ > 0} and then, for z ∈ D we find |h(t, z)| ≤ tΛ−1 .
Therefore, we have α = Λ − 1 and β = 0. On the other hand, for general a, b ∈ C both
end points t = 0 and t = 1 are singular points of g(t). Hence, we are in case 4 of theorem
3.3.1 and, since t1−σ g(t) is bounded as t → 0 for any 0 < σ < 1 we can choose any σ, and
in particular, a σ as close to 1 as we wish. On the other hand, we take γ = max{1, <a}.
Therefore, we can apply theorem 3.3.1 by considering t1 = 1/2 as the only base point of
the standard Taylor expansion of g(t). We have
n−1 k
X 1
g(t) = Ak (a, b) t − + gn (t).
k=0
2
The coefficients Ak (a, b) are elementary functions of a and b. They can be computed as
52 Chapter 4. New uniform expansions of some special functions
In this formula, b(n, k) are certain partial ordinary Bell polynomials [131] that can be
computed recursively by means of
b(0, 0) = 1, b(n, 0) = 0, b(0, k) = 0,
n−k+1
X (n − 1)!
b(n, k) = (−1)j+1 2j b(n − j, k − 1).
(n − j)!
j=1
According to theorem 3.3.1 the remainder terms satisfies Rn (z) = O n1−σ−Λ + n− max{<a,1} ,
with R3 (z) of the order specified above. We can not give more information on the
remainder as we do not have an explicit bound for it due to the difficult expression
for the coefficients An (a, b) of the expansion. The expansion holds uniformly in z with
<z ≥ Λ > 0.
Expansion (4.11) is given in [77, Example 6].
Figure 4.6 and table 4.6 show the accuracy of the uniform expansion (4.11). It is
compared with the well-known power series expansion [103, eqs. 13.2.42 and 13.2.2] and
the asymptotic expansion [103, eq. 13.7.3] of the confluent U function.
2.5 0.4
0.08
0.2
2.0
0.06
2 4 6 8 10
1.5
-0.2
0.04
1.0 -0.4
-0.6
0.02
0.5
-0.8
0 2 4 6 8 10 -1.0 0 2 4 6 8 10
Figure 4.6: Real part (left) and imaginary part (middle) of the approximations of the hyper-
geometric U (a, b, z) function (black, dashed) provided by the power series expansion [103, eqs.
13.2.42 and 13.2.2] (green), the asymptotic expansion [103, eq. 13.7.3] (blue) and the uniform
approximation (4.11) for z ∈ [0, 10eiπ/3 ], a = 0.9 and b = 1.7 when we truncate the series after
n = 4 terms. The figure in the right shows their relative error for a = 1.6, b = 1.9 and real
z ∈ [0, 10] after n = 10 terms of the approximations.
A convergent expansion of E1 (z) uniformly valid in <z > 0 may be derived from
(4.11). But in this particular case we may derive a simpler expansion. To obtain it, we
perform again a logarithmic change of variables defined by u = − log t. We get
1
tz−1
Z
−z
E1 (z) = e dt.
0 1 − log t
1
We identify g(t) = 1−log t
and h(t, z) = tz−1 . We define the region D := {z ∈ C : <z ≥ Λ},
for any Λ > 0. Then, we can take α = Λ − 1 and β = 0. On the other hand, the function
f (t) = tσ−1 g(t) is bounded as t → 0+ , for any σ ∈ (0, 1). Therefore, we can take any
σ ∈ (0, 1) and in particular, we may choose σ as close to 1 as we wish. We also have
γ = 1 as the singular point of g(t) are located at t = 0 (integrable singularity) and at
t = e. Hence, we consider the classical Taylor expansion of g(t) at the point t1 = 1:
n−1
X
g(t) = Ak (1 − t)k + gn (t),
k=0
where (−1)k Ak are the Taylor coefficients of g(t) at t = 1 and gn (t) is the Taylor remain-
der.
The coefficients Ak can be computed recursively using the differential equation satis-
fied by g(t): t(1 − log t)g 0 (t) = g(t), and equating the coefficients of equal powers after
replacing the Taylor expansion for g(t) and g 0 (t). In this way, and after some manipula-
tions we can find the recursive formula [142, eq. 6.10.3]. On the other hand, an explicit
54 Chapter 4. New uniform expansions of some special functions
1.0
0.0010
0.8
0.2 0.0008
0.6
0.0006
2 4 6 8 10
0.4 -0.2
0.0004
-0.4
0.2 0.0002
-0.6
0 2 4 6 8 10 0 5 10 15 20
Figure 4.7: Real part (left) and imaginary part (middle) of the approximations of the function
ez E1 (z) function (black, dashed) provided by the power series expansion [142, eq. 6.6.2] (green),
the asymptotic expansion [142, eq. 6.12.1] (blue) and the uniform approximation (4.14) for
z ∈ [0, 10eiπ/5 ] when we truncate the series after n = 4 terms. The figure in the right shows
their relative error for real z ∈ [0, 20] after n = 10 terms of the approximations.
(α)
where Bn are the Nørlund polynomials [36, eq. 24.16.9]. Formula (4.13) follows from
corollary A.0.2 in appendix A and the derivatives g (n) (0) = (−1)n n!. In particular, the
first coefficients are A0 = 1, A1 = −1, A2 = 1/2, A3 = −1/3, A4 = 1/6, . . . When we
apply theorem 3.3.1 (case 2) to the integral (4.12) we find the expansion1
" n−1 #
X Ak k!
E1 (z) = e−z + Rn (z) . (4.14)
k=0
(z)k+1
For any z with <z ≥ Λ expansion (4.14) is uniformly convergent and the remainder term
is bounded in the form (3.14)
Γ(Λ + σ − 1)Γ(n + 1)
|Rn (z)| ≤ M ,
Γ(n + Λ + σ)
for a certain M > 0 independent of n and z. Therefore, Rn (z) = O n1−σ−Λ with σ as
close to 1 as we wish.
Expansion (4.14) may be found in [142, eq. 6.10.1] or [145, Ch. 17, §3].
In figure 4.7 we compare the accuracy of the well-known power series [142, eq. 6.6.2]
and asymptotic expansion [142, eq. 6.12.1] of the function ez E1 (z) with the also well-
known factorial series (4.14) that holds uniformly in <z ≥ Λ > 0. The convergent
features are exhibited in table 4.7.
We consider the symmetric elliptic integrals RF (x, y, z) and RD (x, y, z). These func-
tions were introduced by Carlson who showed that they are more appropriate for numer-
ical purposes than the standard elliptic integrals [23, 24, 25, 26, 27]. They are defined
by the following integral representations [28, eq. 19.16.1], [28, eq. 19.16.5]
1 ∞
Z
ds
RF (x, y, z) := √ √ √ , (4.15)
2 0 s+x s+y s+z
3 ∞
Z
ds
RD (x, y, z) := √ √ p , (4.16)
2 0 s + x s + y (s + z)3
where, in both integrals, x, y, z ∈ C \ (−∞, 0] except that one or more of x, y, z may be 0
when the corresponding integral converges. We assume that the three variables x, y and
z are different, because otherwise the functions RF and RD are elementary functions.
Moreover, for the sake of simplicity and due to the symmetry of the variables we also
assume that one of the variables of RF , say z, is positive. Similarly, we assume for RD
that either x, y ∈ C \ (−∞, 0] and z > 0 or y, z ∈ C \ (−∞, 0] and x > 0. The results
that we are going to derive can be extended to complex values of the positive variables
by using analytic continuation arguments (see Appendix B).
The functions RF and RD are homogeneous functions of degree, respectively, −1/2
and −3/2 in its three variables. Moreover, the function RF is symmetric in its three
variables whereas only the variables x and y play the same role in the integrand of the
function RD . Therefore, instead of the function RF (x, y, z) we consider the function
F (x, y) defined in (4.17) below. On the other hand, instead of considering the function
RD (x, y, z), we consider, if z > 0 the function G1 (x, y) given in (4.18) or, if x > 0, the
function G2 (y, z) defined in (4.19). These functions are defined by means of
√ 1 ∞
Z
ds
F (x, y) := zRF (z(1 + x), z(1 + y), z) = √ √ √ , (4.17)
2 0 s+x+1 s+y+1 s+1
with x, y ∈ C \ (−∞, −1].
√ 3
Z ∞
ds
G1 (x, y) := z 3 RD (z(1 + x), z(1 + y), z) = √ √ p ,
2 0 s+x+1 s+y+1 (s + 1)3
(4.18)
with x, y ∈ C \ (−∞, −1].
√ 3
Z ∞
ds
G2 (y, z) := x3 RD (x, x(1 + y), x(1 + z)) = √ √ p ,
2 0 s + 1 s + y + 1 (s + z + 1)3
(4.19)
y, z ∈ C \ (−∞, −1].
56 Chapter 4. New uniform expansions of some special functions
The functions F, G1 and G2 are functions of only two variables and they are more
convenient than the functions RF and RD to derive new convergent expansions uniformly
valid when one of its variables runs in a large set of the complex plane. And all the results
that we are going to obtain can be translated to the symmetric elliptic functions by means
of the connection formulas
1 x−z y−z
RF (x, y, z) = √ F , , x, y ∈ C \ (−∞, 0], z > 0. (4.20)
z z z
1 x−z y−z
RD (x, y, z) = √ G1 , , x, y ∈ C \ (−∞, 0], z > 0. (4.21)
z3 z z
1 y−x z−x
RD (x, y, z) = √ G2 , , y, z ∈ C \ (−∞, 0], x > 0. (4.22)
x3 x x
Nevertheless, the integrals (4.17), (4.18) and (4.19) are not suitable for the derivation
of uniform expansions, as the interval of integration is unbounded. In the general theory of
uniform expansion developed in chapter 3 a logarithmic change of variables was indicated
to obtain a compact interval of integration. That change of variables was useful from a
theoretical point of view and in some examples 4.6, 4.7, but it would be too difficult to
work with it if we introduce it in any of the formulas (4.17), (4.18) or (4.19). Instead,
we introduce a simpler change of variables s 7→ t given by 1 + s = 1/t. In this way, we
obtain
1 1
Z
dt
F (x, y) = √√ √ , x, y ∈ C \ (−∞, −1], (4.23)
2 0 t 1 + xt 1 + yt
√
3 1
Z
tdt
G1 (x, y) = √ √ , x, y ∈ C \ (−∞, −1] (4.24)
2 0 1 + xt 1 + yt
and √
Z 1
3 tdt
G2 (y, z) = √ p , y, z ∈ C \ (−∞, −1]. (4.25)
2 0 1 + yt (1 + zt)3
For the sake of convenience and generality, we consider the more general function
Z 1
tc dt
F(a, b, c; x, y) := a b
, x, y ∈ C \ (−∞, −1] (4.26)
0 (1 + xt) (1 + yt)
Lemma 4.8.1. For any fixed angle θ ∈ [π/2, π), consider the extended sector
[
S(θ) := {y ∈ C : | arg(y)| ≤ θ}
\ \
{y ∈ C : | arg(y)| > θ} {y ∈ C : |y + 1| > sin θ} {y ∈ C : |y + 1/2| ≤ 1/2} ,
(4.27)
4.8 The symmetric elliptic integrals RD (x, y, z) and RF (x, y, z) 57
Im(y)
R3
P+ R2
π-θ R1
θ
-1 sin
-1/2 Re(y)
P-
Figure 4.8: The region S(θ) defined in equation (4.27) is marked in green. Its boundary is the
portion of the rays arg y = ±θ exterior to the open disk R2 (the disk of center −1/2 and radius
1/2) and also the portion of the circle |y + 1| = sin θ interior to R2 . In the limit case θ → π the
region S(θ) becomes the cutted complex plane C \ (−∞, −1].
that is depicted in green in figure 4.8. Then, for any y ∈ S(θ), t ∈ [0, 1] and <b ≥ 0 the
function h(t, y) := tc (1 + yt)−b can be uniformly bounded in the form
|h(t, y)| ≤ t<c eπ|=b| (sin θ)−<b . (4.28)
Proof. By splitting b into its real and imaginary parts, it is straightforward to check that
|h(t, y)| = |tc | |(1 + yt)−b | ≤ |(1 + yt)−<b | t<c eπ|=b| . (4.29)
To bound the first factor, we divide the region C \ (−∞, −1] in three different parts
named R1 , R2 and R3 that are shown in figure 4.8. In each region, we search the absolute
maximum of the function |(1 + yt)−<b | for t ∈ [0, 1] (absolute that depends on the value
of <y). We have:
For y ∈ R1 := {y ∈ C : <y ≥ 0}, the maximum is attained at t = 0 and its value
is 1.
If y ∈ R2 := {y ∈ C : |y + 1/2| < 1/2} = {y ∈ C : <y < −|y|2 }, the maximum is
attained at t = 1 and its value is |1 + y|−<b .
When y ∈ R3 := {y ∈ C \ (−∞, −1] : −|y|2 ≤ <y < 0}, the maximum is attained
at t = −<y/|y|2 and its value is | sin(arg y)|−<b . Note that the region R3 is the
left half complex plane <y < 0 with both, the straight (−∞, −1] and the disk R2 ,
removed.
For any fixed angle θ ∈ [π/2, π), the rays arg y = ±θ cut the boundary of the disk R2
at the points P± = (− cos2 θ, ± sin θ cos θ). On the one hand, at the portions of the rays
arg y = ±θ that are inside the region R3 (that is a portion of the boundary of S(θ)) we
have that
| sin(arg y)|−<b ≤ (sin θ)−<b .
On the other hand, at the portion of the circle |y + 1| = sin θ that is inside the disk R2 ,
that is, the remaining portion of the boundary of S(θ), we have that
|1 + y|−<b ≤ (sin θ)−<b .
58 Chapter 4. New uniform expansions of some special functions
In any case, the last two inequalities together with (4.29) prove the uniform bound (4.28)
given in the lemma.
Remark 4.8.2. In the limit case θ → π, the region S(θ) becomes the cutted complex plane
C \ (−∞, −1].
Lemma 4.8.3. For any x ∈ C \ (−∞, −1], consider a point of the complex plane w(x)
given by any of the following three formulas:
1 if | arg(x + 1)| < π/2,
1 <(x+1)−|x+1|
w(x) = × 1 + i =(x+1) if 0 < | arg(x + 1)| < π, (4.30)
2
0 if |x| < 1.
Then,
|xw(x)| < |1 + xw(x)| , |x (1 − w(x))| < |1 + xw(x)| . (4.31)
Moreover, for arg(x+1) = π, the two inequalities |xw| < |1+xw| and |x(1−w)| < |1+xw|
can not be simultaneously satisfied for any value of w.
Proof. In order to simplify the notation, we simply write w(x) = w. If we divide the
second inequality of (4.31) by |x + 1|, then (4.31) becomes
x x
|xw| < |1 + xw| , (w − 1) < 1 + (w − 1) ,
x+1 x+1
x
which mean that the distance from both points, xw and x+1 (w −1), to the point −1 must
be larger than the distance to the point 0. This statement is equivalent to the following
two inequalities:
1 x 1
<(xw) > − , < (w − 1) > − .
2 x+1 2
We define θ := arg(x+1), r := |x+1| and we choose w to be of the form w = (a+1/2)+ib
with a, b ∈ R. Then, the above inequalities read
1 1
cos θ − a − b sin θ + cos θ > 0,
r 2 (4.32)
1
(r − cos θ) a − b sin θ + cos θ > 0.
2
If we take θ = π, the inequalities in (4.32) become
1 r
<a<− .
2(r + 1) 2(r + 1)
Therefore, when arg(x + 1) = π, inequalities (4.31) do not hold for any w ∈ C. On the
other side, we have that:
For |θ| < π/2 the two inequalities in (4.32) hold for a = 0 = b, that is, w = 1/2
(first line in (4.30)).
For θ 6= 0, π the two inequalities in (4.32) hold for a = 0 and b < cot(θ)/2 if
0 < θ < π or b > cot(θ)/2 if −π < θ < 0. In particular, the inequalities hold for
the value of w given in the second line of (4.30).
4.8 The symmetric elliptic integrals RD (x, y, z) and RF (x, y, z) 59
For r < 2 cos θ the two inequalities in (4.32) hold for a = −1/2 and b = 0, that is,
for w = 0. The condition r < 2 cos θ is equivalent to |x| < 1 (last line in (4.30)).
Remark 4.8.4. Note that, for many values of x, there is an overlaping in the sector that
defines w(x) in equation (4.30), that is, there are values of x for which several values
of w(x) are possible. This is not important and lemma 4.8.3 states that for any possible
choice, inequalities (4.31) hold.
With the help of the lemmas proved above we are in conditions for obtaining a uni-
formly convergent expansion of the function F introduced in (4.26) and defined by the
integral representation
Z 1
tc dt
F(a, b, c; x, y) := a b
, x, y ∈ C \ (−∞, −1] (4.33)
0 (1 + xt) (1 + yt)
With this election of the base point and with the help of lemma 4.8.3, we have that
x(w(x)−t)
1+xw(x)
< 1 for any t ∈ [0, 1]. Thence, expansion (4.34) is convergent and the remainder
rn (t; x, w, a) can be written in the form
∞
1 X (a)k (−x)k (t − w(x))k
rn (t; x, w, a) =
(1 + xw)a k=n k! (1 + xw(x))k
n
(a)n x (w(x) − t) 1, n + a x (w(x) − t)
= 2 F1 .
(1 + xw(x))a n! 1 + xw(x) n+1 1 + xw(x)
For t ∈ [0, 1], it holds that |t − w(x)| ≤ |w(x)| and then, using the series definition of the
Gauss hypergeometric function [104, eq. 15.2.1], we find that
n
|(a)n | xw(x) 1, n + <a xw(x)
|rn (t; x, w, a)| ≤ 2 F1 . (4.35)
|(1 + xw(x))a | n! 1 + xw(x) n+1 1 + xw(x)
On the other hand, for any fixed angle θ ∈ [π/2, π) we have that, using lemma 4.8.1,
the function h(t, y) can be bounded in the form
tc
|h(t, y)| := ≤ t<c eπ|=b| | sin θ|−<b := H(t), (4.36)
(1 + yt)b
for any y ∈ S(θ), being S(θ) ⊂ C \ (−∞, −1] the region defined in lemma 4.8.1 and
depicted in figure 4.8. Note that H(t) is an integrable function on [0, 1].
60 Chapter 4. New uniform expansions of some special functions
We replace g(t) in the integral (4.33) by the right hand side of (4.34) and, interchang-
ing summation and integration, we obtain
n−1 k
1 X (a)k −x
F(a, b, c; x, y) = Ak (b, c; y, w(x)) + Rn (a, b, c; x, y),
[1 + xw(x)]a k=0 k! 1 + xw(x)
(4.37)
where the coefficients Ak (b, c; y, w) are related to the moments of the function h(t, y) in
the form
Z 1 Z 1
Ak (b, c; y, w) := k
(t − w) h(t, y)dt = (t − w)k tc (1 + yt)−b dt. (4.38)
0 0
Taking into account the bounds (4.36) and (4.35) and integrating, we find the bound
Remark 4.8.5. Due to the election of w(x) given by lemma 4.8.3, the branch point t =
−1/x of the function g(t) = (1 + xt)−a is located outside the disk of convergence of
the Taylor expansion of g(t) at t = w(x). Moreover, the integration interval [0, 1] is
completely contained in that disk of convergence and then, according to case 1 of theorem
3.3.1 the convergence of expansion (4.37) should be of exponential type as shown by (4.41)
xw(x)
(recall that, by lemma 4.8.3, xw(x)+1 < 1.)
4.8 The symmetric elliptic integrals RD (x, y, z) and RF (x, y, z) 61
Remark 4.8.6. We have considered above the point t = w(x) as base point for the Taylor
expansion of the function g(t). This election is valid for any value of x. However, if we
impose the more demanding restriction |x| < 1 (see the third line of (4.30) in lemma
4.8.3) we could consider t = 0 as the base point for the Taylor expansion of g(t). Then,
following the same steps as above we would obtain a simpler expansion of F(a, b, c; x, y)
with a more accurate error bound (see [21, Theorem 3.2]).
Recall that the function F(a, b, c; x, y) generalizes the functions F (x, y), G1 (x, y) and
G2 (y, z) given in (4.17), (4.18) and (4.19) respectively and related with the symmet-
ric elliptic integral by means of the connection formulas (4.20), (4.21) and (4.22). We
have that F (x, y) = 12 F (1/2, 1/2, −1/2; x, y), G1 (x, y) = 32 F (1/2, 1/2, 1/2; x, y) and
G2 (y, z) = 23 F (1/2, 3/2, 1/2; y, z). Hence, a convergent expansion for those functions,
that holds uniformly for y ∈ S(θ) follows from (4.37).
In general, the coefficients Ak (b, c; x, y) given in the right hand side of approximation
(4.37) are given in (4.39) in terms of Gauss hypergeometric 2 F1 functions. That is, ex-
pansion (4.37) is not given in terms of elementary functions. However, in the particular
cases when F represents any one of the functions F , G1 or G2 related with the ellip-
tic symmetric integrals, then the corresponding coefficients Ak (b, c; x, y) are elementary
functions of x and y. More precisely, from (4.39) we have
k
k (−w)k−j
X b, j + c + 1
Ak (b, c; y, w) = 2 F1 −y .
j=0
j c+j+1 j+c+2
If b = 1/2 and c = −1/2 (that is, we are considering the coefficients of the function
F (x, y)) we have
Z 1 k−1/2
1/2, k + 1/2 t
2 F1 − y := √ dt
k + 3/2 0 1 + yt
k
2 (−1)k Γ(k + 1/2) √ p X (−y)−j (1/2 − k)j−1
= √ k+1/2 arcsinh( y) − 1 + y .
πy k! j=1
k − j + 1 (−k) j−1
(4.42)
If b = 1/2 and c = 1/2 (we are considering the coefficients of the function G1 (x, y))
we have
Z 1 k+1/2
1/2, k + 3/2 t
2 F1 − y := √ dt
k + 5/2 0 1 + yt
√ k
! (4.43)
(3/2)k arcsinh( y) p X j!(−y j )
= √ − 1+y .
(k + 1)!(−y)k+1 y j=0
(3/2)j
If b = 3/2 and c = 1/2 (we are considering the coefficients of the function G2 (y, z))
we have
Z 1
tk+1/2
3/2, k + 3/2
2 F1 − z := p dt
k + 5/2 0 (1 + zt)3
" √ k j
!# (4.44)
(3/2)k 2arcsinh( z) 1 X (j − 1)!(−z)
= √ + √ −2 + .
k!(−z)k z3 z 1+z j=1
(3/2)j
62 Chapter 4. New uniform expansions of some special functions
0.010 0.010
1.2
1.1
0.008 0.008
1.0
0.006 0.006
0.9
0.7
0.002 0.002
0.6
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
Figure 4.9: The figure on the left shows the function RF (1, y, 2) (black, dashed) and the
approximations provided by the uniform approximation (4.37) using the connection formula
(4.20) (red), the series expansion for small values of y [65, Corollary 3.1, eq. 3.1] (green) and
the series expansion for large values of y [65, Corollary 3.1, eq. 3.1] (blue) after n = 2 terms
of the approximations. The other two pictures show the relative error for x = 1, z = 3 and
real y ∈ [0, 10] after n = 4 terms on the approximations of RF (x, y, z) (middle) and RD (x, y, z)
(right) using the uniform approximation (4.37) and the connection formula (4.20) or (4.21)
(red), the series expansion for small values of y [65, Corollary 3.1, eq. 3.1] or [65, Corollary
3.4, eq. 3.14] (green) and the series expansion for large values of y [65, Corollary 3.1, eq. 3.1]
or [65, Corollary 3.4, eq. 3.14] (green) (blue). We have taken w = 1/2 as base point for the
uniform approximations.
Remark 4.8.7. To obtain the integral representations (4.17) and (4.18) for the functions
F (x, y) and G1 (x, y) we have assumed that z > 0. Then, if we apply any of the connection
formulas (4.20) or (4.21) to obtain a uniformly convergent expansion of RF or RD from
(4.37) we are restricted, in principle, to z > 0. However, in appendix B it is shown by
means of analytical continuation arguments that the integral representations (4.17) and
(4.18), and therefore the corresponding approximations, are valid in the bigger sector
√ √
(x + 4y) arcsin y x y + 1
F (x, y) = − + ε(x),
4y 3/2 4y
with |ε(x)| ≤ 0.075x2 ≤ 0.075, valid for 0 ≤ x < 1 and uniformly in y, with <y > 0.
4.9 The Gauss hypergeometric and the incomplete beta functions 63
valid for t ∈ [0, 1] and −z ∈ S(θ), with S(θ) defined in (4.27) and depicted in figure 4.8.
Therefore, we find α = 0 and β = 0.
Thus, using theorem 3.3.1 we obtain the convergent expansion [43, eq. 2.2], uniformly
for −z ∈ S(θ). The remainder of the expansion after n terms satisfies Rn (z, a, b, c) =
O n−b + n−(c−b) , as n → ∞ [43, eq. 2.9].
As an illustration of the kind of approximation [43, eq. 2.2] that we find, we have,
for example
1 3 √
,
2 2 384 − 1120z + 2100z 2 + 525z 3 + 1 − z (−384 + 928z − 1684z 2 − 1275z 3 )
2 F1 5
z = √
2 560 2z 4
+ ε(z)
valid for <z ≤ 0 with |ε(z)| < 0.0784. For z = 0 the right hand side of this formula must
be understood in the limit sense. The above formula follows from [43, eq. 2.2] after 3
terms. The bound for the remainder ε(z) follows from [43, eq. 2.19].
64 Chapter 4. New uniform expansions of some special functions
0.05
0.65
0.04
1.0
0.45
0.03
0.8
0.25
0.02
0.6
0.05
-10 -5 5 10 0.01
0.4 -0.15
-10 -5 0 5 10 -10 -8 -6 -4 -2 0
Figure 4.10: Real part (left) and imaginary part (middle) of the approximations of the Gauss
hypergeometric 2 F1 (a, b, c; z) function (black, dashed) provided by the power series definition
[104, eq. 15.2.1] (green), the asymptotic expansion [104, eqs. 15.2.1 and 15.8.2] (blue) and the
uniform expansion [43, eq. 2.2] (red), for z ∈ [−10eiπ/4 , 10eiπ/4 ] after n = 4 terms, for a = 0.5,
b = 1.3 and c = 2.5. The figure on the right shows their relative errors for n = 10, a = 0.8,
b = 1.7, c = 3.4 and real z ∈ [−10, 1]. The behavior is similar for other values of z and a, b, c.
In figure 4.10 and table 4.9 we show the uniform and asymptotic properties of ex-
pansion [43, eq. 2.2]. We also compare it with the well-known power series definition
[104, eq. 15.2.1] and asymptotic expansion [104, eqs. 15.2.1 and 15.8.2] for large z of the
Gauss hypergeometric function.
Moreover, for <b ≤ 1 and −z ∈ S(θ) we can obtain a uniformly convergent expansion
of the incomplete beta function Bz (a, b). In particular, we may derive the expansion [44,
eq. 2.2]. The remainder of this expansion after n terms satisfies Rn (z, a, b) = O (n−a ),
as n → ∞ [44, eq. 2.7].
Remark 4.9.1. For integers values of b and c, the hypergeometric 2 F1 function is a rational
function of a and z. In this case, the uniform expansion [43, eq. 2.2] after n terms is
exact if n is large enough (n ≥ c) as the bound [43, eq. 2.18] for the remainder vanishes.
the point t = 0 and t = 1 are singular points of g(t). Then, we are again in case 4 of
theorem 3.3.1 with σ = c − b and γ = b.
We need to bound the function h(t, z) uniformly in a large set of the complex z-plane.
To this end, we define, for any r ∈ (0, 1], the region
Remark 4.9.2. For integers values of b and c, the hypergeometric 2 F1 function is a rational
function of a and z. In this case, the uniform expansion [43, eq. 2.2] after n terms is
exact if n is large enough (n ≥ c) as the bound [43, eq. 2.9] for the remainder vanishes.
66 Chapter 4. New uniform expansions of some special functions
1.0 0.010
0.9 0.3
0.008
0.8 0.2
0.6
-10 -5 5 10 0.004
0.5
-0.1
0.4 0.002
-0.2
0.3
-0.3
-10 -5 0 5 10 -10 -8 -6 -4 -2 0
Figure 4.11: Real part (left) and imaginary part (middle) of the approximations of the incom-
plete beta function z −a (1−z)1−b Bz (a, b) (black, dashed) provided by the power series definition
[124] (green), the asymptotic expansion [116, eq. 8.17.7]+[104, eqs. 15.2.1 and 15.8.2] (blue)
and the uniform expansion [44, eq. 3.2] (red) for z ∈ [−10eiπ/6 , 10eiπ/6 ] after n = 4 terms, for
a = 1.5 + 0.75i and b = 2.25 + 0.25i. The figure on the right shows their relative errors for
n = 10, a = 1.1, b = 2.2 and real z ∈ [−10, 1]. The behavior is similar for other values of z and
a, b.
Chapter 5
An Application of the Uniform
Expansions: A Series Representation
of the Volterra Function
67
68 Chapter 5. An application of the uniform expansions
with <β > −1 and t > 0. After certain manipulations and using a uniform approximation
of a factor in the integrand, we derive a convergent expansion of µ(t, β, α) in terms of
incomplete gamma functions. The results of this chapter are based on [70].
5.1 Introduction
The Volterra function µ(t, β, α) is defined by the integral representation (5.1), but there
are some particular notations that are usually adopted in some particular cases, namely:
Z ∞ Z ∞
tu du tu+α du
ν(t) := µ(t, 0, 0) = , ν(t, α) := µ(t, 0, α) = ,
0 Γ(u + 1) 0 Γ(u + α + 1)
Z ∞ u β
1 t u du
µ(t, β) := µ(t, β, 0) = .
Γ(β + 1) 0 Γ(u + 1)
The function µ(t, β, α) and any of its particular cases, ν(t), ν(t, α) or µ(t, β) are
analytic functions of the variable t with branch-points at t = 0 and ∞ and no more
singularities [40, Ch. 18, §18.3]; and ν(t, α) and µ(t, β, α) are entire functions of α.
Moreover, the variable β is restricted so that <β > −1 in order for the integral (5.1) to
be convergent, but the definition of µ(t, β, α) can be extended to the whole β−plane by
repeated integration by parts [40, Ch. 18, §18.3, eq. 4]
Z ∞ Z ∞
tu+α uβ tu+α
1 −1 β+1 d
µ(t, β, α) : = du = u du
Γ(β + 1) 0 Γ(u + α + 1) Γ(β + 2) 0 dt Γ(u + α + 1)
Z ∞
(−1)m tu+α
β+m d
= ... = u du
Γ(β + m + 1) 0 dt Γ(u + α + m)
for <β > −m − 1, with m ∈ N ∪ {0}. The so extended function µ(t, β, α) is also an entire
function of β.
These functions are named after the italian mathematician Vito Volterra who, in
1916, introduced them as solution to certain integral equations with a logarithmic kernel
[148]. Since then, these functions have played an important role in differents fields in
mathematics. Specially important are their applications in integral equations and opera-
tional calculus, as shown by many french mathematicians in the forties of the last century
[30, 31, 53, 121, 122, 123]; and due to integral formulas such as [40, Ch. 18, §18.3, eq.
22]
Z ∞ 2
t
exp − µ(t, β, α)dt = 2β+1 y 1/2 π 1/2 µ(y, β, α/2) <α > −1, <y > 0,
0 4y
that relate the Volterra function with a certain integral transform of itself.
For this reason, Volterra functions take a central role in the theory of the Laplace
transform as they are the Laplace transform of simple functions. For example [40, Ch.
18, §18.3, eq. 17],
Z ∞
tβ e−st
dt = eαs µ e−s , β, α ,
<β > −1,
0 Γ(α + t + 1)
and Z ∞
e−st µ (t, β, α) dt = s−α−1 (log s)−β−1 , <α > −1, <s > 1.
0
5.1 Introduction 69
Besides, the function ν(t, α) appears in the formulaR∞of Paley-Wiener for the inversion of
a Laplace transformation [112]. That is, if f (s) = 0 e−st F (t)dt, then
Z ∞
1
F (t) = lim f (s) [ν (st, −1/2 + λi) − ν (st, −1/2 − λi)] ds.
λ→∞ 2πi 0
More recently, Mainardi et al. [49, 85, 86] have studied the Volterra functions in con-
nection with solutions to fractional relaxation/diffusion equations of distributed order;
and Apelblat [2] has collected a comprehensive set of information about this function,
providing a historical perspective, abundant bibliography, important identities and sev-
eral integral transforms.
On the other hand, the so called Fransén-Robinson constant, denoted by F , and
defined by menas of Z ∞
1
F := dx,
0 Γ(x)
is a special case of the Volterra function as F = µ(1, 0, −1) = µ(1, 1, 0). This constant
has interesting applications in statistics [48]: Imagine a certain probability model whose
density function decreases faster than e−cx , for any positive constant c. Then, the recip-
rocal gamma function may be used as density function and the value of the constant F
is needed for the sake of normalization.
In spite of the importance that the Volterra function has in many branches of math-
ematics, most books on special function do not consider it. For this reason, asymptotic
expansions have not yet been fully investigated, although asymptotic expansions as t → 0
and t → ∞ may be found, respectively, in [40] and [155]. The results are summarized in
[49] and given below:
Define the coefficients
(−1)n 1 dn
(α) 1
Dn := µ (1, −n − 1, α) = .
n! n! dxn Γ(α + x + 1) x=0
Then,
∞ −β−1−n
α
X 1
µ(t, β, α) ∼ t (β + 1)n Dn(α) log , as t → 0, with t ∈ C \ [1, +∞).
n=0
t
(5.2)
On the other hand,
where H(t, β, α) is the expansion in the right hand side of (5.2) and
∞ (α,β)
X En
E(t, β, α) := et tβ−n ,
n=0
Γ(β + 1 − n)
(α,β)
being En the coefficients given by
(−1)n dn (1 − x)−α−1 (−x)β+1
En(α,β) := .
n! dxn logβ+1 (1 − x) x=0
Nevertheless, convergent expansions are not available in the literature. The aim of
this chapter of the thesis is to fill this gap by deriving a family of convergent series of
70 Chapter 5. An application of the uniform expansions
Figure 5.1: A possible path C is obtained by joining a circle of radius R > |t| with the two
straight lines =(w) = ±ε, for any 0 < ε < R, traversed in the counterclockwise direction.
the Volterra function. To obtain it, in section 5.2 we derive an integral representation
of the Volterra function different from the one given as definition in (5.1) and that is
more suitable for our analysis. Then, a uniform expansion of a factor of the integrand in
the new integral representation is used to obtain a convergent expansion of the Volterra
function (section 5.3). The resulting expansion is given in terms of a family of definite
integrals that are related to the incomplete gamma function. Finally, some numerical
experiments are shown in section 5.4.
Figure 5.2: The integration contour L (left) in the integral (5.5) crosses the real line at a
certain c > 0. It can be deformed to the rectangle-like contour (middle) consisting of a vertical
line at <s = c and two straight lines at height =s = ±iπ/2 that go up to infinity parallel to
the real axis. This contour can be further deformed to a similar rectangle Γ, but with height
=s = ±π (right). All three paths are traversed in the clockwise direction.
The contour C is, for example, the one depicted in figure 5.1 and made up of two straight
lines and a circle centered at w = 0 with radius R > |t|. If <α ≥ −1 we can deform the
contour C to the vertical line C 0 := {R + iu : −∞ < u < +∞}, with R > |t|. Then
∞
eR tα eiu (R + iu)−α−1
Z
µ(t, β, α) = du, R > |t|.
2π −∞ logβ+1 [(R + iu) /t]
with the only restriction <b < 0 if c simultaneously satisfies <c = 0 and =c ∈ [−π, π].
Lemma 5.3.1. The function φ(a, b, c) defined in (5.9) admits, for <c > 0 and <b > −1,
the following integral representation:
Z ∞ b −cx
2e−ca x e sin [π(a + x)]
φ(a, b, c) = dx. (5.10)
Γ(b + 1) 0 a+x
Proof. We consider the well-known integral representation of the Γ function [4, eq. 5.9.1]
given by Z ∞
Γ(ν)
= e−zx xν−1 dx, <ν > 0, <z > 0. (5.11)
zν 0
Thus, we can write
Z ∞
1 1
= xb e−(c+iu)x dx, <b > −1, <c > 0,
(c + iu)b+1 Γ(b + 1) 0
which, replaced into (5.9) and using Fubini’s theorem to interchange the order of inte-
gration yields
Z ∞ Z π
1 −a(c+iu) b −(c+iu)x
φ(a, b, c) = e xe du dx
Γ(b + 1) 0 −π
Z ∞ Z π
e−ca
b −cx −i(a+x)u
= xe e du dx.
Γ(b + 1) 0 −π
2 sin[π(a+x)]
The inner integral can be straightforwardly computed and equals a+x
, which gives
(5.10) and proves the lemma.
2. If a 6= 0 and b ∈ (N ∪ {0}),
(−1)b [log (a(c + iπ)) − log(c + iπ)]
b
φ(a, b, c) = ia Γ(−b, a(c + iπ)) + −
b!
(5.13)
(−1)b [log (a(c − iπ)) − log(c − iπ)]
−Γ(−b, a(c − iπ)) − .
b!
3. If a = 0 and b 6= 0,
2 sin [b arctan(π/c)]
φ(0, b, c) = √ b . (5.14)
b c2 + π 2
4. If a = 0 and b = 0, π
φ(0, 0, c) = 2 arctan . (5.15)
c
Proof. Formulas (5.14) and (5.15) follow by directly computing the integral (5.9) (with
a = 0). For example, if b 6= 0,
Z π " #
b b
du i i (c − iπ) − (c + iπ)
= (c + iπ)−b − (c − iπ)−b =
φ(0, b, c) = .
−π (c + iu)
b+1 b b (c2 + π 2 )b
As
√ b √ b
(c + iπ)b = c2 + π 2 eib arctan(π/c) and (c − iπ)b = c2 + π 2 e−ib arctan(π/c) ,
it follows that
√ b
(c − iπ)b − (c + iπ)b = −2i c2 + π 2 sin [b arctan(π/c)]
and we find (5.14). In a similar way, or taking the limit value b → 0 in (5.14), formula
(5.15) can be found.
In order to prove the remaining formulas (5.12) and (5.13), we consider the series
expansion of the exponential function in the integral (5.9) to find
Z π −a(c+iu) Z πX ∞
e (−a)n
φ(a, b, c) = b+1
du = (c + iu)n−b−1 du. (5.16)
−π (c + iu) −π n=0 n!
We invoke to the dominated convergence theorem to interchange summation and inte-
gration. Then, we have to distinguish two cases depending on whether b = 0, 1, 2, . . ., or
not.
If b ∈
/ (N ∪ {0}), we get
∞
X (−a)n i (c − iπ)n−b − (c + iπ)n−b
φ(a, b, c) =
n=0
n! n−b
∞ ∞
1 X [−a(c + iπ)]n 1 X [−a(c − iπ)]n
= − .
i(c + iπ)b n=0 n!(n − b) i(c − iπ)b n=0 n!(n − b)
From the series representation [116, eq. 8.7.1] of the regularized incomplete gamma
function γ ∗ (α, z) we have that
Γ(−b) Γ(−b)
φ(a, b, c) = b
γ ∗ (−b, a(c + iπ)) − γ ∗ (−b, a(c − iπ))
i(c + iπ) i(c − iπ)b
and formula (5.12) follows.
74 Chapter 5. An application of the uniform expansions
Then,
Z ∞ Z ∞
e−iαπ
−αu −teu −(u+iπ)v β
µ∞ (t, β, α, c) = e e e v dv du−
2πiΓ(β + 1) c 0
Z ∞ Z ∞
eiαπ
−αu −teu −(u−iπ)v β
− e e e v dv du.
2πiΓ(β + 1) c 0
where, in the last equality, we have used that the last integral above with respect to u is
in fact an incomplete gamma function, as can be easily seen by performing the change
5.3 Series representations of the Volterra function 75
of variables u 7→ w given by w = teu and using the integral definition of the incomplete
gamma function [116, eq. 8.2.1].
It might seem that representation (5.17) of the function µ∞ (t, β, α, c) is less conve-
nient than formula (5.8) to derive a convergent expansion of the function µ∞ (t, β, α, c).
However, as we will see below, they key point is to use the uniformly convergent expan-
sion of the incomplete gamma function Γ(a, z) valid for all <a < 0 given in section 4.4.
That expansion is uniformly valid in the whole interval of integration and then, if we
replace the gamma function Γ(−v − α, tec ) in the integrand of (5.17) by its uniformly
convergent expansion and we interchange summation and integration, we hope to obtain
an expansion of the funcion µ∞ (t, β, α, c) that is convergent.
In the following theorem we show that the above idea is not only formal, but rigorous,
by finding accurate bounds for the remainder of the expansion.
Theorem 5.3.3. Let α, β ∈ C with <α > 0 and <β > −1, and t, c ∈ R with t > 0
and c > 0. Then, for any n ∈ N, the function µ∞ (t, β, α, c) defined in (5.8) admits the
representation
c n−1 k
e−te X X k c(j+1)
µ∞ (t, β, α, c) = c
Lk (te ) e φ (α + j + 1, β, c)+Rn∞ (t, c, β, α), (5.18)
2π k=0 j=0
j
where the functions φ(a, b, c) have been defined in (5.9) and computed in terms of incom-
plete gamma functions in lemma 5.3.2, and Ln (x) are the Laguerre polynomials.
The remainder Rn∞ (t, c, β, α) is bounded in the form
c /2−c<α+π|=α|
e−te Γ (<β + 1) Γ (<α) 1
|Rn∞ (t, c, β, α)| ≤ . (5.19)
π|Γ(β + 1)|c<β+1 n<α
Moreover, if <β ≥ −1/2 we have the simpler bound
c
e−te /2−c<α+π|=α| Γ (<α) 1
|Rn∞ (t, c, β, α)| ≤ . (5.20)
π sech(π=β)c<β+1 n<α
p
In any case, expansion (5.18) is convergent with a convergence rate of power type.
That is
Rn∞ (t, c, β, α) = O n−<α ,
as n → ∞.
Proof. We consider the integral representation of the function µ∞ (t, c, β, α) given in the
last line of (5.17) and we replace the function Γ(−v − α, tec ) by its convergent expan-
sion (4.9) given in corollary 4.4.1 of section 4.4 and uniformly valid for v in the whole
integration interval (0, ∞). We get
Z ∞
−1
µ∞ (t, β, α, c) = v β sin [(v + α)π] tv+α Γ(−v − α, tec ) dv
πΓ(β + 1) 0
Z ∞ " n−1
−1 β v+α −tec c −v−α
X
= v sin [(v + α)π] t e (te ) Lk (tec )B (1 + v + α, k + 1) +
πΓ(β + 1) 0 k=0
n
#
X
+ Rn (−v − α, tec ) dv
k=0
c n−1 Z ∞
−e−te e−αc X
= c
Lk (te ) v β e−vc sin [(v + α)π] B (1 + v + α, k + 1) dv + Rn∞ (t, c, β, α),
πΓ(β + 1) k=0 0
(5.21)
76 Chapter 5. An application of the uniform expansions
k
(5.23)
(−1)j
X k
= .
j=0
j v + α + j + 1
where the last equality follows from lemma 5.3.1. In this manner we get the right hand
side of expansion (5.18).
It remains to show that the remainder Rn∞ (t, β, α, c) can be bounded as given in the
theorem. Using the first inequality of (4.10) in corollary 4.4.1 we have
c
|tv+α Rn (−v − α, tec )| ≤ e−te e−c<α e−cv B (v + <α, n + 1) .
Now, an easy computation shows that the above beta function is a decreasing function of
v. Thus, for all v ∈ [0, ∞), B (v + <α, n + 1) ≤ B (<α, n + 1), and using the inequality
[4, eq. 5.6.8] as we did in the last line of the proof of lemma 4.4.1 we find
c c
|tv+α Rn (−v − α, tec )| ≤ e−te e−c<α e−cv B (<α, n + 1) ≤ e−te e−c<α e−cv Γ(<α)n−<α ,
Then, from (5.22), the previous bound for tv+α Rn (−v−α, tec ) and the bound | sin [π(v + α)] | ≤
eπ|=α| , we find
c
e−te e−c<α Γ(<α)n−<α ∞ −cv <β
Z
∞
|Rn (t, c, β, α)| ≤ e v dv.
π|Γ(β + 1)| 0
Remark 5.3.4. In principle, for a fixed c > 0, the bounds (5.19) or (5.20) in theorem
5.3.3 show that the expansion (5.18) is convergent, but the order of convergence is only
of power type: O n−<α , as n → ∞. However, if we let c depende on n in the form
tec = λn, for any fixed positive paramater λ, that is, if we take c = log (λn/t), then
(5.20) becomes
eπ|=α|−λn/2 t<α Γ(<α)
|Rn∞ (t, c, β, α)| ≤ <β+1 2<α
sech(π=β) n ,
p
πλ<α log λn
t
where the functions φ(a, b, c) are defined in (5.9) and computed in lemma 5.3.2 in terms
of incomplete gamma functions. The remainder Rn0 (t, c, β, α) can be bounded in the form
c
e−c<α+π(|=α|+|=β|/2)+te (tec )n
|Rn0 (t, c, β, α)|
≤ . (5.25)
c<β+1 n!
Therefore, Rn0 (t, c, β, α) = O (tec+1 )n n−1/2−n as n → ∞ and expansion (5.24) is con-
vergent.
Proof. We recall the definition of µ0 (t, β, α, c), given in (5.7) by
c iu
e−αc π e−iαu ete e
Z
µ0 (t, β, α, c) := du.
2π −π (c + iu)β+1
The exponential is an entire function whose well-known series expansion is
n−1 k
tec+iu
X t
e = ek(c+iu) + rn0 (t, c, u), for any n ∈ N,
k=0
k!
with ∞
X (tec )k c γ(n, tec )
|rn0 (t, c, u)| ≤ = ete . (5.26)
k=n
k! Γ(n)
We can replace the exponential function in the integral definition of µ0 (t, β, α, c) to get
n−1 k k(c+iu)
!
e−αc π e−iαu
Z X t e
µ0 (t, β, α, c) = + rn0 (t, c, u) du
2π −π (c + iu)β+1 k=0 k!
n−1
1 X tk π e−(α−k)(c+iu)
Z
= du + Rn0 (t, c, β, α).
2π k=0 k! −π (c + iu)β+1
78 Chapter 5. An application of the uniform expansions
Comparing with the definition of the φ(a, b, c), (5.9), we find the right hand side of (5.24)
with
e−αc π e−iαu
Z
0
Rn (t, c, β, α) := r0 (t, c, u)du.
2π −π (c + iu)β+1 n
In order to find the bound (5.25) we use the bound obtained in (5.26) for the remainder
rn0 (t, c, u). We immediately find
c π
e−c <α ete γ(n, tec ) π|=α|
Z
du
|Rn0 (t, c, β, α)| ≤ e .
2π Γ(n) −π |(c + iu)β+1 |
Now, we have |(c+iu)β+1 | ≥ c<β+1 e|=β|π/2 and γ(n, tec ) ≤ (tec )n /n which follows from
the integral definition of the incomplete gamma function [116, eq. 8.2.1]. Accordingly
we find the bound (5.25).
Finally, the asymptotic behavior of the remainder Rn0 (t, c, β, α) as n → ∞ is found
after an application of the Stirling formula for the factorial n!. Then, expression (5.24)
is a convergent expansion of the function µ0 (t, β, α, c), for any c > 0.
Remark 5.3.6. The asymptotic behavior of the remainder Rn0 (t, c, β, α) shows that expan-
sion (5.24) is convergent. However, as we did in remark 5.3.4, we do not take a fixed
parameter c > 0, but we let c depend on n in the form tec = λn, for a fixed parameter
λ > 0. Then, using the Stirling approximation of the gamma function and the asymptotic
behavior of the incomplete gamma function γ(n, λn) given in [116, eq. 8.11.6] (valid for
0 < λ < 1) in (5.25), we find
<α π(|=α|+|=β|/2)
√
t e 2π (eλ)n
|Rn0 (t, log(λn/t), β, α) | ≤ <β+1/2 n<α+1/2
,
(1 − λ)λ<α log λn
t
(eλ)n
Rn0 (t, log(λn/t), β, α) = O , as n → ∞.
(log n)<β+1 n<α+1/2
In particular, for any number 0 < λ < e−1 , we can take c = log(λn/t) and the rate of
convergence of expansion (5.24) is exponential.
Corollary 5.3.7. Let α, β ∈ C with <α > 0, <β > −1 and t > 0. Then, for any positive
integer n and any arbitrary positive number c > 0, the Volterra function µ(t, β, α) satisfies
n−1
" k
1 X −tec c
X k c(j+1)
µ(t, β, α) = e Lk (te ) e φ(α + j + 1, β, c)
2π k=0 j=0
j
# (5.27)
n−1
X tk
+ φ(α − k, β, c) + Rn (t, c, β, α),
k=0
k!
5.3 Series representations of the Volterra function 79
where Lk (x) are the Laguerre polynomials and φ(a, b, c) are the functions defined in (5.9)
that may be written in terms of incomplete gamma functions (see lemma 5.3.2). The
remainder Rn (t, c, β, α) is bounded in the form
c c
e−c<α+π|=α| e−te /2 M (β)Γ(<α) eπ|=β|/2+te γ(n, tec )
|Rn (t, c, β, α)| ≤ + , (5.28)
c<β+1 π n<α Γ(n)
Proof. The results follows from the relation (5.6) and theorems 5.3.3 and 5.3.5, as well
as the remarks 5.3.4 and 5.3.6. The value λ0 corresponds to the only value of λ such that
0 < λ < e−1 that makes the terms inside the brackets in (5.29) to behave similarly. That
λn
is, if we impose the condition e− 2 = (eλ)n , we get the equation log(eλ) + λ2 = 0, whose
only solution is λ = λ0 := 0.31436990296762807 . . .. That is, for λ = λ0 , the expansions
(5.18) of µ∞ (t, β, α, c) and (5.24) of µ0 (t, β, α, c) have a similar convergence rate, which
is exponential.
Remark 5.3.8. If we set c = log λn , with 0 < λ < e−1 , we may need to take a large
t
number of terms to obtain a valid result when using expansion (5.27), as c > 0 if and
only if n ≥ λt . In any case, as the expansion is convergent, by taking enough terms we
can find as many precision as required.
Remark 5.3.9. The expansion (5.27) derived for the Volterra function in corollary 5.3.7 is
not given in terms of elementary functions [132] but rather in terms of incomplete gamma
functions. Nevertheless, the expansion is useful as incomplete gamma functions are sim-
pler than the Volterra function. On the one hand, the Volterra function depends on three
parameters α, β and t whereas the incomplete gamma function only depends on two. On
the other hand, incomplete gamma functions have been investigated in detail and there is
an extensive literature about this funcion. Apart from its integral definition, a lot of infor-
mation is known about incomplete gamma functions [116]: more integral representations,
special values, convergent series expansions, asymptotic expansions in several domains of
the variables (standard and uniform), recurrence relations and derivatives, relations to
other special functions, continued fractions, inequalities, zeros, integrals, sums... On the
other side, as it has been pointed out in the introduction, the Volterra function has barely
been investigated, probably due to the gamma function in the denominator of its integral
definition (5.1).
80 Chapter 5. An application of the uniform expansions
Remark 5.3.10. The Volterra function µ(t, β, α) is real for real variables. On the other
hand, in approximation (5.27) we can find the function φ(a, b, c) for certain parameters
a, b and c, which are computed in lemma 5.3.2 in terms of incomplete gamma functions
with complex arguments. This should not lead to confusion, the functions φ(a, b, c) are
also real for real values of its parameters, as it is shown from its integral representation
given in lemma 5.3.1.
Remark 5.3.11. The expansion derived in corollary 5.3.7 is only valid for <α > 0, whereas
the Volterra function is defined for any α ∈ C. For negative values of <α we may use
the recurrence relation [2]
which can be derived from the integral definition (5.1) of the Volterra function and the
recurrence relation Γ(z + 1) = zΓ(z) for the gamma function.
105
1000
10
0.100
0.001
10-5
2 4 6 8
Figure 5.3: Plot, in logarithmic scale, of the Volterra function µ(t, 3.1, 2.5) (black, dashed), the
first term of the asymptotic expansion for small t (5.2) (blue), the 4 first terms of the asymptotic
expansion for large t (5.3) (green) and the 12 first terms of the expansion given by (5.27) with
c = 0.1 (red). The asymptotic expansion (5.2) is to be preferred for small t and the asymptotic
expansion (5.3) is to be preferred for large t, whereas the convergent expansion (5.27) given in
corollay 5.3.7 performs better for intermediate values and produces an approximation that is
in general more satisfactory.
t n = 10 n = 15 n = 20 n = 25
0.3 8.98 · 10−2 3.16 · 10−2 2.00 · 10−2 6.89 · 10−3
0.7 4.80 · 10−3 7.43 · 10−4 4.96 · 10−4 3.51 · 10−5
1 4.01 · 10−4 1.91 · 10−4 5.57 · 10−5 3.44 · 10−5
1.3 8.82 · 10−4 2.86 · 10−5 1.50 · 10−5 6.46 · 10−6
1.7 3.98 · 10−3 9.12 · 10−6 3.53 · 10−6 4.54 · 10−7
2 9.53 · 10−3 4.33 · 10−6 9.41 · 10−7 5.50 · 10−7
Table 5.1: Relative error provided by formula (5.27) in corollary 5.3.7 for α = 2.4 + 1.1i,
β = 0.8 + 2.9i, c = 0.5 and several values of t.
104
10
0.01
10-5 2 4 6 8
Figure 5.4: Plot, in logarithmic scale, of the Volterra function µ(t, 3.1, 2.5) (black), the asymp-
totic approximations for small t (5.2) (continuous lines), the asymptotic expansions for large t
(5.3) (dot-dashed lines) and the 12 first terms of the expansion (5.27) given in corollary 5.3.7
with c = 0.1 (red, dashed). In the asymptotic formulas (5.2), (5.3) for small and large t we have
taken several terms of the approximation: n = 1 (dark yellow), n = 3 (orange), n = 6 (green),
n = 9 (blue) and n = 12 (purple).
82 Chapter 5. An application of the uniform expansions
t n = 30 n = 40 n = 50 n = 60
4 8.06 · 10−3 8.52 · 10−9 8.00 · 10−11 7.84 · 10−11
6 8.46 · 10−2 4.36 · 10−6 7.96 · 10−12 3.47 · 10−12
8 4.80 · 10−1 2.11 · 10−4 1.49 · 10−9 5.50 · 10−12
10 − 2.88 · 10−3 1.57 · 10−7 4.18 · 10−11
12 − 2.09 · 10+9 5.37 · 10−6 1.18 · 10−9
15 − − 2.67 · 10+7 2.42 · 10−6
Table 5.2: Relative error provided by formula (5.27) in corollary 5.3.7 for α = 14, β = 17,
c = log(λ0 n/t) and several values of t.
β n = 10 n = 15 n = 20 n = 25
(3 + 2i) · 10−8 5.82 · 10−3 2.90 · 10−5 5.15 · 10−8 3.27 · 10−8
0.5 + 1.2i 1.18 · 10−2 5.77 · 10−5 8.00 · 10−8 3.13 · 10−8
4.6 + 3.7i 6.17 · 10−3 2.72 · 10−5 3.07 · 10−8 1.43 · 10−9
8.1 + 6.9i 4.46 · 10−1 9.81 · 10−5 8.33 · 10−8 4.29 · 10−10
Table 5.3: Relative error provided by formula (5.27) in corollary 5.3.7 for α = 2.3 + 0.7i,
t = 1.2, c = 1.4 and several values of β.
β n = 30 n = 40 n = 50 n = 60
5 1.34 · 10−4 5.79 · 10−12 9.99 · 10−13 7.22 · 10−13
15 1.43 · 10−1 2.69 · 10−6 3.54 · 10−11 3.19 · 10−11
30 1.41 2.19 · 10−2 3.15 · 10−7 5.84 · 10−10
50 8.00 · 10+9 4.27 7.15 · 10−3 4.20 · 10−6
Table 5.4: Relative error provided by formula (5.27) in corollary 5.3.7 for α = 18, t = 5,
c = log(λ0 n/t) and several values of β.
α n = 10 n = 15 n = 20 n = 25
(6 + 7i) · 10−5 1.31 · 10−4 2.60 · 10−7 1.12 · 10−7 4.50 · 10−9
1.4 + 0.2i 4.70 · 10−4 2.98 · 10−7 9.36 · 10−9 1.00 · 10−9
2 + 3.2i 7.28 · 10−1 1.22 · 10−3 5.33 · 10−6 9.08 · 10−7
3.6 + 4.1i 2.28 · 10+1 3.64 · 10−2 1.93 · 10−5 9.02 · 10−7
Table 5.5: Relative error provided by formula (5.27) in corollary 5.3.7 for β = 4.9 + 2.7i,
t = 0.8, c = 1.6 and several values of α.
α n = 30 n = 40 n = 50 n = 60
10 1.61 · 10−2 1.43 · 10−8 3.44 · 10−11 1.51 · 10−11
20 3.48 9.93 · 10−4 2.26 · 10−9 6.90 · 10−10
30 1.61 · 10+1 3.39 7.23 · 10−5 8.40 · 10−8
40 7.76 · 10+7 2.03 · 10+2 2.69 3.91 · 10−2
Table 5.6: Relative error provided by formula (5.27) in corollary 5.3.7 for β = 19.7 + 11.4i,
t = 4, c = log(λ0 n/t) and several values of α.
Chapter 6
A Dual Vision of Uniform Expansions:
New Expansions of the Generalized
Hypergeometric p+1Fp Function
that hold in a large, unbounded region D of the complex plane containing a selected point,
say z = 0. The basic idea to obtain this expansion was simple: to consider a multi-point
Taylor expansion of the function g(t) in m selected points 0 ≤ t1 < t2 < . . . < tm ≤ 1.
That expansion is uniformly and absolutely convergent in a certain lemniscate Dr . It
avoids all singular points of the function g(t) and, if the base point are cleverly chosen, the
integration interval (0, 1) is contained in the lemniscate Dr . Moreover, as the function g(t)
is independent of the variable z, so is its multi-point Taylor remainder rn (t). Therefore,
when we replace g(t) by its multi-point Taylor series and interchange summation and
integration, we obtain a convergent expansion for the integral F (z). If, in addition, the
R 1 bounded independently of z in a certain region D, the resulting
function h(t, z) can be
remainder Rn (z) = 0 h(t, z)rn (t)dt can also be bounded independently of z and, as a
result, the expansion for the integral transform F (z) is valid in a large region D of the
complex plane.
On the other hand, if we consider an expansion for the factor h(t, z) we usually obtain
an expansion for the integral transform F (z) that is valid in a small region. Typically,
if h(t, z) = h(tz), the classical power series expansion of F (z) is found by considering
the Taylor series expansion of h(tz) at t = 0. In general, the function h(t, z) has some
finite singularities and then the radius of its Taylor series is small. As a consequence,
the power series of many integral transforms (6.1) converge in a small disk. On the
other hand, asymptotic expansions are found by considering an expansion of g(t) at
the asymptotically relevant point of h(t, z). In general, these expansions are a powerful
numerical tool and valid for large values of the variable, but they are divergent. Therefore,
it is necessary to study the optimal truncation term of the expansion and also to find
accurate bounds for the remainder. In this chapter we investigate a new approach that
83
84 Chapter 6. A dual vision of uniform expansions
The approach taken in [76, 79, 83] is the following, that we illustrate with the Gauss
hypergeometric 2 F1 function: Consider its integral representation
Z 1
a, b Γ(c)
2 F1 z = tb−1 (1 − t)c−b−1 (1 − zt)−a dt,
c Γ(b)Γ(c − b) 0
valid for <c > <b > 0 and | arg(1 − z)| < π. The standard Taylor series expansion of
the 2 F1 function can be found by considering the Taylor series expansion at t = 0 of
f (t) := (1 − zt)−a and interchanging summation and integration. The validity of that
expansion (the disk |z| < 1) is determined by the following two requirements:
(i) The interval of integration (0, 1) must be completely contained in the domain of
convergence of the series expansion of f (t), which is a disk of center t = 0 and
radius r ≥ 1, say Dr .
(ii) The branch point t = 1/z of f (t) must be located outside the domain Dr . In
other words, z must be located in a region Sr = the inverse to the exterior of Dr :
Sr = (DrEXT )−1 = {z ∈ C : |z| < r−1 }.
Therefore, the smaller Dr is, the bigger Sr would be. As Dr must verify (i), the smallest
possible value of r is 1 + ε, with ε > 0 small, and then Sr = {z ∈ C : |z| < 1}. But, using
multi-point Taylor expansions, the region Dr becomes thinner and one expects that the
corresponding multi-point region Sr becomes bigger. Following this idea, in [76, 79, 83]
6.1 The generalized hypergeometric p+1 Fp function 85
multi-point Taylor expansions of the function f (t) are considered with requirements (i)
and (ii) adapted to the lemniscate of convergence of the multi-point Taylor expansion.
This procedure is used to obtain new series expansions of the 2 F1 , 3 F2 and M hyper-
geometric functions that hold in large regions of the complex plane. Those regions are
unbounded and contain small values of the variable.
We can generalize this idea to any given integral transform F (z) of the form (6.1)
by considering the multi-point Taylor series expansion of the function h(t, z) at certain,
cleverly selected base points 0 ≤ t1 < t2 < . . . < tm ≤ 1. That expansion is uniformly
and absolutely convergent in a certain lemniscate Om . Then, we choose the m points
satisfying the following two requirements:
(I) The integration interval (0, 1) in (6.1) is contained in Om .
(II) The variable z runs in a certain region S such that the singularities of h(t, z) as a
function of t, say t0 (z) (in general, they depend on z), are not contained in Om .
Depending on the location of the singularities of h(t, z) the region S may be very large,
possibly unbounded and containing small values of the variable z. This method can be
viewed as a dual version of the uniform expansions introduced in chapter 3.
We illustrate this idea with the generalized hypergeometric p+1 Fp (a, ~b, ~c; z) function.
In particular, we generalize the results given in [76, 83] for the 2 F1 and 3 F2 functions
to an arbitrary number of parameters 2p + 1. Moreover, we not only consider the Euler
form of the integral representation of the generalized hypergeometric function, but a more
general integral (6.4) given below and derived in appendix C.
The results of this chapter have been published in [72].
where the symbol ∗ indicates that the entry ak + 1 − ak is omitted. Thus, using (6.2)
and (6.3) the function p+1 Fp (a, ~b, ~c; z) can be computed in the exterior of the unit disk
|z| ≥ ρ > 1, with ρ depending on numerical requirements.
Other expansions for the hypergeometric function p+1 Fp , in terms of elementary func-
tions, may be found in the literature. For example, Nørlund [100] found an expansion in
z
powers of z−1 [3, eq. 16.10.2] that converges in the half plane <z < 1/2, Bühring [18]
obtained an expansion in inverse powers of z − z0 [3, eq. 16.8.9], for arbitrary complex z0 ,
valid in the region |z − z0 | > max{|z0 |, |z0 − 1|} and | arg(z0 − z)| < π, and in [55, §4.1]
we can find an inverse factorial series representation of the generalized hypergeometric
p+1 Fp function that converges in the half-plane <z < 0.
All those expansions, the power series defintion (6.2), the Taylor expansion at z = ∞
(6.2)-(6.3), Nørlund’s expansion and Bühring’s expansion are given in terms of elementary
functions but, for general values of (a, ~b, ~c), none of them is convergent in unbounded
regions of the complex plane containing the indented closed unit disk D∗ := {z ∈ C :
|z| ≤ 1, z 6= 1}. Therefore, we seek new convergent expansions of p+1 Fp given in terms
of elementary functions that are convergent in large regions of the complex plane and, in
particular, in unbounded regions containing the indented closed unit disk D∗ .
The starting point is the following integral representation of the generalized hyperge-
ometric function p+1 Fp [72, eq. 5] that we derive in appendix C
p
!
a, ~b Y Z
bs −1 cs −bs −1
p+1 Fp z = A(bs , cs ) ts (1 − ts ) dts (1 − T z)−a , (6.4)
~c s=1 L
For any possibility in the choice of L and A(bs , cs ) in the integral (6.4), we consider the
function f (T ) := (1 − zT )−a . When we replace f (T ) in (6.4) by its standard Taylor series
expansion at T = 0, interchange summation and integration and use the corresponding
integral representation of the beta function given in [4, §5.12] we obtain the power series
definition (6.2). If |z| < 1 , the Taylor series expansion of f (T ) at T = 0 converges
uniformly in z for any (t1 , . . . , tp ) ∈ L × . . . × L in any of the four possible integration
domains considered in (6.4). Therefore, expansion (6.2) is convergent in the unit disk
|z| < 1.
Following [76, 83] and pointed out in the introduction of this chapter, the region of
convergence |z| < 1 of expansion (6.2) is determined by requiremets (i) and (ii) above.
Therefore, the smaller Dr is, the bigger the convergence region of the expansion of p+1 Fp
would be. For this reason, we consider multi-point Taylor expansions as the lemniscate of
6.2 New expansions for the ~
p+1 Fp (a, b, ~
c; z) function 87
(i) Appropriately selecting the base points of the multi-point Taylor expansion of f (T ),
the domain X (that is squeezed as much as possible around the real interval [0, 1])
is contained in D, the domain of convergence of the multi-point Taylor expansion
of f (T ).
As we will see, the use of multi-point Taylor expansions result in expansions that are
covergent in large (possibly unbounded) regions of the complex plane that contain the
indented closed unit disk D∗ .
Theorem 6.2.1. For arbitrary w ∈ C define W := max{|w|, |1 − w|} and the region
S := {z ∈ C : |1 − wz| > |z|W }. Then, for any z ∈ S and (a, ~b, ~c) ∈ Λ,
! n−1 k ! !
a, ~b ~ ~
X (a) k wz −k, b 1 a,b
p+1 Fp z = (1 − wz)−a p+1 Fp + Rn z .
~c k=0
k! wz − 1 ~
c w ~
c
(6.5)
On the one hand, the functions p+1 Fp (−k, ~b, ~c; 1/w) are polynomials in the variable 1/w
that can be computed using lemma D.0.1 in Appendix D.
On the other hand, the remainder term Rn (a, ~b, ~c; z) is uniformly bounded in compact
sets of S in the form
!
n
a,~b zW
Rn z ≤ M (a, ~b, ~c; z) , (6.6)
~c 1 − wz
with M (a, ~b, ~c; z) > 0 independent of n. Furthermore, if <ck > <bk > 0 for all k =
1, 2, . . . , p, then M (a, ~b, ~c; z) is also independent of ~b and ~c. In this case
!
a,~b eπ|=a| Γ(n + <a)|zW |n
Rn z ≤ . (6.7)
~c |Γ(a)|n!(|1 − wz| − |zW |)n+<a
Consequently
!
a,~b |zW |n n<a−1
Rn z =O , as n → ∞. (6.8)
~c (|1 − wz| − |zW |)n+<a
88 Chapter 6. A dual vision of uniform expansions
Proof. Consider the Taylor series expansion of f (T ) = (1 − zT )−a at the generic point
T = w ∈ C:
n−1 k
−a
X (a)k z(T − w)
f (T ) = (1 − wz) + rn (T ). (6.9)
k=0
k! 1 − wz
Define D = {T ∈ C : |T − w| < W + ε with ε > 0 small}. Then, condition (i) is satisfied,
that is, D contains the domain X. For z ∈ S condition (ii) is satisfied too, that is, if
z ∈ S then 1/z ∈ / D.
Moreover, for T ∈ [0, 1] we have |T −w| ≤ W . Therefore, expansion (6.9) is convergent
for all z ∈ S and the remainder rn (T ) satisfies
∞ k
−a
X (a)k z(T − w)
rn (T ) := (1 − wz)
k=n
k! 1 − wz
n (6.10)
(a)n z(T − w) 1, n + a z(T − w)
= (1 − wz)−a 2 F1 .
n! 1 − wz n+1 1 − wz
For z ∈ S we can replace f (T ) in (6.4) by the right hand side of (6.9). After
interchanging summation and integration we find
! n−1 k
a, ~b
−a
X (a)k z
p+1 Fp z = (1 − wz)
~c k=0
k! wz − 1
p
!
Y Z
a,~b
× A(bs , cs ) tbss −1 (1 − ts )cs −bs −1 dts (w − T )k + Rn z .
s=1 L ~
c
(6.11)
Writing (w − T )k = wk (1 − T /w)k and using the integral representation (6.4) we find,
in any of the four possibilities for the path L and the constants A(bs , cs ), the right hand
side of expansion (6.5). The remainder Rn (a, ~b, ~c; z) of the expansion is given by
p
!
a,~b Y Z
bs −1 cs −bs −1
Rn z = A(bs , cs ) ts (1 − ts ) dts rn (T ), (6.12)
~c s=1 L
w=i
1+i
w= 2
Figure 6.1: The minimal domain of convergence of the standard Taylor expansion of f (T ) at
T = w containing X is a disk of center w and radius r = W + ε (figures left). The region S
(figures right) is the inverse of the exterior of D: it is a disk for <w < 1/2 whereas for <w ≥ 1/2
it is a half-plane.
Introducing this bound into (6.12) and using the corresponding integral representation
of the beta function [4, §5.12] to integrate we obtain (6.6).
On the other hand, from (6.10) and using the integral representation of the hyperge-
ometric 2 F1 function [104, eq. 15.6.1] we find the bound
n
−a |(a)n | zW 1, n + <a zW
|rn (T )| ≤ |(1 − wz) | 2 F1 (6.16)
n! 1 − wz n+1 1 − wz
In the case that, for all k = 1, 2, . . . , p, <ck > <bk > 0 we can consider Euler’s form
for the path L, that is, L = [0, 1], and introducing bound (6.16) into (6.12) and after
some easy computations we find (6.7). The asymptotic behavior as n → ∞ (6.8) follows
from (6.7) and Stirling’s formula for the gamma function.
Remark 6.2.2. The domain of convergence of the expansion (6.5) is the region S := {z ∈
C : |1 − wz| > |z|W }. If <w < 1/2 the region S is a disk of center w̄/(2<w − 1) and
radius (|w − 1|)/(1 − 2<w). For <w ≥ 1/2 the region S is the half-plane <(wz) < 1/2
(see figure 6.1).
Remark 6.2.3. When w = 1 expansion (6.5) is Nørlund’s expansion [3, eq. 16.10.2] given
in terms of powers of z/(z − 1). It converges in the half-plane <z < 1/2.
where, for n = 1, 2, . . ., the coefficients A2n (a, z) and Bn2 (a, z) and p+1 Fp (−n, ~b, ~c; 1) are
rational functions of a and z and ~b and ~c respectively. They can be computed recursively
using lemmas D.0.2 and D.0.1 in Appendix D.
On the other hand, the remainder RN (a, ~b, ~c; z) is bounded in the forn
!
N
a,~b ~ (1/2 − q)2 z 2
RN z ≤ M (a, b, ~c; z) , (6.19)
~c (1 − qz) [1 + (q − 1)z]
with M (a, ~b, ~c; z) > 0 independent of N . Moreover, if <ck > <bk > 0, for all k =
1, 2, . . . , p, M (a, ~b, ~c; z) is also independent of ~b and ~c.
Proof. We consider the two-point Taylor expansion of f (T ) = (1 − zT )−a at T = q and
T =1−q
N
X −1
An (a, z) + Bn2 (a, z)T [(T − q)(T + q − 1)]n + rN (T ),
2
f (T ) = (6.20)
n=0
valid in a Cassini oval Dq with foci at T = q and T = 1 − q and a certain radius r > 0
to be determined, that is, Dq = {T ∈ C : |(T − q)(T + q − 1)| < r}. In formula (6.20)
rN (T ) is the two-point Taylor remainder after N terms.
An explicit formula for the coefficients A2n (a, z) and Bn2 (a, z) can be derived from for-
mulas (2.2)-(2.3) and also using the recurrent algorithm described after (2.8) in chapter 2.
Instead, the recurrence relation (D.3) given in Appendix D is found from the differential
equation satisfied by the function f (T ), namely, (1 − zT )f 0 (T ) = azf (T ).
We determine the radius r of the Cassini oval Dq as follows: The interval [0, 1] is
contained in Dq whenever
where the Taylor remainder rN (T ) can be written in terms of a Cauchy integral (2.4).
Analogously to the proof in theorem 6.2.1, the remainder rN (T ) can be bounded in the
form
N
(T − q)(T + q − 1)
|rN (T )| ≤ C(a, z) 1 1 , (6.22)
z
− q z
+ q − 1 − ε
with C(a, z) > 0 independent of T and N . Moreover, when T ∈ X the absolute value of
the numerator above is bounded by (1/2 − q)2 and, as (6.22) is valid for arbitrarily small
ε > 0, we find
N
(1/2 − q)2 z 2
|rN (T )| ≤ C(a, z) (6.23)
(1 − qz) [1 + (q − 1)z]
Introducing this bound into (6.21) we obtain (6.19). If in addition <ck > <bk > 0 for
all k = 1, 2, . . . , p we can use the Euler form of (6.4) (when L = [0, 1]) as we did in the
proof of theorem 6.2.1 to show that M (a, ~b, ~c; z) is independent of ~b and ~c.
Remark 6.2.5. Two particular cases are to be highlighted. On the one hand, when we take
T = 0 and T = 1 as base points for the two-point Taylor expansion of f (T ) = (1 − zT )−a ,
that is, for q = 0, expansion (6.18) is the simplest one. In particular, the inner sum that
runs from k = 0 to k = n in (6.18) contains only one non-zero term and formulas
√ (D.2),
(D.3) and (D.4) become easier too. On the other √ hand, for q = q :=
0√ (2 − 2)/4 the base
points for the Taylor expansions are T = (2− 2)/4 and T = (2+ 2)/4. In this case the
region Sq20 is the largest possible one. In both cases, q = 0 and q = q0 the indented closed
unit disk D∗ = {z ∈ C : |z| ≤ 1, z 6= 1} is containded in the region Sq2 of convergence.
Moreover, for q = q0 the region Sq20 is unbounded (see figure 6.2).
Figure 6.2: The Cassini oval Dq of convergence of the two-point Taylor expansion of f (T ) with
foci at T = q, T = 1 − q and radius r (left) contains the interval [0, 1] and avoids the branch
point T = 1/z for r > ( 21 −q)2 and q ∈ [0, q0 ]. The region Sq2 (right) is the inverse of the exterior
of the disk Dq . For q = 0 (top) the region of convergence S02 is bounded. For q = q0 (bottom),
the region Sq20 is the largest possible one and it is unbounded. We see that the thinner Dq is
the larger Sq2 becomes.
! N −1 X
n p
" −−−→ !
a, ~b
X n Y (b )
s k b+k
p+1 Fp z= (−2)k−n A3n (a, z)Hnq −−−→
~c n=0 k=0
k s=1
(cs )k c+k
p −−−−−−→ ! p −−−−−−→ !#
bs + k 3 q b+k+1 (bs + k)(bs + k + 1) 3 b+k+2
Y Y
+ Bn (a, z)Hn −−−−−−→ + Cn (a, z)Hnq −−−−−−→
c
s=1 s
+ k c+k+1 s=1
(c s + k)(c s + k + 1) c+k+2
!
a,~b
+ RN z ,
~c
(6.25)
with
~b
! n p −−−→ !
X n Y (bs )m −m, b + m
Hnq := [q(1 − q)]n−m (−1)m p+1 Fp −−−→ 1 . (6.26)
~c m=0
m s=1
(cs )m c+m
The coefficients A3n (a, z), Bn3 (a, z) and Cn3 (a, z) and the functions Hnq (~b, ~c) are, respec-
tively, rational functions of a and z and ~b and ~c. They can be computed recursively using
lemmas D.0.3 and D.0.1 in Appendix D.
6.2 New expansions for the ~
p+1 Fp (a, b, ~
c; z) function 93
with M (a, ~b, ~c; z) > 0 independent of N . Moreover, if <ck > <bk > 0 for all k =
1, 2, . . . , p, then M (a, ~b, ~c; z) is also independent of ~b and ~c.
Proof. The proof follows the same steps as the proof of theorems 6.2.1 and 6.2.4, but
considering the three-point Taylor expansion of f (T ) = (1 − zT )−a at the base points
T = q, T = 1/2 and T = 1 − q, with q ∈ [0, 1/2). We have
N
X −1
A3n (a, z) + Bn3 (a, z)T + Cn3 (a, z)T 2 [(T − q)(T − 1/2)(T + q − 1)]n + rN (T ),
f (T ) =
n=0
(6.28)
valid in a lemniscate Dq of foci T = q, T = 1/2 and T = 1 − q and a certain radius r > 0,
that is, Dq = {T ∈ C : |(T − q)(T − 1/2)(T + q − 1)| < r}. To determine de radius r, we
require that the lemniscate Dq satisfy requirements (i) and (ii).
On the one hand, the interval [0, 1] is contained in Dq if
q(1 − q) (1 − 2q)3
r ≥ sup {|(T − q)(T − 1/2)(T − 1 + q)|} = max , √ .
T ∈[0,1] 2 12 3
√ 3
q(1−q)
Let q1 := 2− 3
4
be the solution of the equation 2
= (1−2q)
√ .
12 3
Then, for 0 ≤ q ≤ q1 ,
(1−2q)3
expansion (6.28) satisfies condition (i) for r ≥ √ . On the
other hand, it satisfies
12 3
condition (ii) if 1/z ∈
/ Dq , that is, for any
1 1 1 1
r< −q − +q−1 .
z z 2 z
3
The smallest possible r is r = (1−2q)
√
12 3
+ ε, for ε > 0 small and then, the largest possible
region of convergence S is the one given in (6.24): Sq3 .
For z ∈ Sq3 we introduce the expansion (6.28) into the integral (6.4), we interchange
summation and integration and we obtain (6.25), with
p
!
a,~b Y Z
bs −1 cs −bs −1
RN z := A(bs , cs ) ts (1 − ts ) dts rN (T ). (6.29)
~c s=1 L
From Cauchy’s formula for the Taylor remainder rN (T ) we find the bound
N
(T − q)(T − 1/2)(T + q − 1)
|rN (T )| ≤ C(a, z) 1
,
− q z1 − 21 z1 + q − 1 − ε
z
with C(a, z) > 0 independent of T and N . This bound is valid for arbitrarily small ε√> 0
and, when T ∈ X, we also have |(T − q)(T − 1/2)(T + q − 1)| ≤ (1 − 2q)3 /(12 3).
Therefore
N
(1 − 2q)3 z 3
|rN (T )| ≤ C(a, z) √ .
6 3 (1 − qz) (2 − z) [1 + (q − 1)z]
94 Chapter 6. A dual vision of uniform expansions
Figure 6.3: The minimal domain of convergence Dq of the three-point Taylor expansion of f√
(T )
at the base point T = q, T = 1/2 and T = 1 − q is√a lemniscate of radius r > (1 − 2q)3 /(12 3)
(figures left with q = 0 (top) and q = q1 := (2 − 3)/4 (bottom)). The corresponding region
Sq3 , the exterior of the inverse of Dq , is shown on the right for q = 0 (top) and q = q1 (bottom).
Introducing this bound into (6.29) and integrating we find (6.27). If in addition <ck >
<bk > 0 for all k = 1, 2, . . . , p, the Euler form of (6.4) can be used to show that
M (a, ~b, ~c; z) is independent of ~b and ~c, as we did in the proof of theorem 6.2.1.
√
Remark 6.2.7. We highlight the two extremal cases q = 0 and q = q1 := (2 − 3)/4. In
the former case, expansion (6.25) becomes the simplest one whereas in the latter case the
region Sq31 of convergence is the largest possible one. Both regions contain the indented
closed unit disk D∗ := {z ∈ C : |z| ≤ 1, z 6= 1}. Moreover, the region Sq31 is unbounded
(see figure 6.3).
integrals we have seen that the expansions are also valid for any (a, ~b, ~c) ∈ Λ, that is, for
any complex value of the parameters a, ~b and ~c except 1 − ck ∈ N.
We have considered the 2− and 3−point Taylor expansion of f (T ) at the generic
base points (q, 1 − q) and (q, 1/2, 1 − q) with q ∈ [0, 1/2) in contrast to [76, 83] where
only the case q = 0 is studied. In this manner, we have been able to find large regions
of convergence for the new expansions of the hypergeometric functions. Most of these
regions Sq contain the indented closed unit disk D∗ and some of them are unbounded (see
figures 6.1, 6.2, 6.3). We could also have considered the possibility of expanding f (T )
at four (or even more) points located along the interval [0, 1]. In this way, the region of
convergence S would be larger than the regions Sq2 and Sq3 shown in figures 6.2 and 6.3.
However, the integrals defining the coefficients of the expansion as well as the recurrence
relations for the coefficients An (a, z), Bn (a, z), Cn (a, z), . . . (see Appendix D), become
more and more complicated.
Furthermore, the derived expansions (6.5), (6.18) and (6.25) are given in terms of
rational functions of z and the parameters bk and ck . We have obtained bounds for the re-
mainder of the expansions that show the convergence of the expansions in the correspond-
ing regions Sq with an exponential rate of convergence: RN (a, ~b, ~c; z) = O A(z)−N , as
N → ∞, for a certain A(z) > 1 that is, in each case, related with the convergence region
Sq . We have only given a precise bound for the remainder in subsection 6.2.1 when we
used the classical Taylor expansion of the function f (T ). In the other cases when we used
multi-point Taylor expansion, we have given the bound for the remainder in terms of a
certain constant M (a, ~b, ~c; z). The reasons for this are the following: (i) we do not have
at our disposal an appropriate expression for the remainder rN (T ) of the two- and three-
point Taylor exansions of f (T ) that would lead to the derivation of an accurate constant
C(a, z); and (ii) the loop integrals are less appropriate than the Euler form when we want
to find an explicit and accurate value of the bounding constant M (a, ~b, ~c; z). And Euler’s
form is only valid in the case when <ck > <bk > 0, for all k ∈ 1, 2, . . . , p.
In tables 6.1-6.6 we show some numerical experiments about the accuracy of the ex-
pansions (6.5), (6.18) and (6.25) to evaluate the functions 4 F3 (a, ~b, ~c; z) and 8 F7 (a, ~b, ~c; z)
for some values of the parameters a, ~b and ~c and different values of the variable z. In
expansion (6.5) we have taken w = 1/2 as base point for the Taylor expansion. On the
other hand, in expansions (6.18) and (6.25) we have taken the simplest case q = 0 and
the value of q that maximizes the size of the region of convergence, that is, q = q0 for
(6.18) and q = q1 for (6.25). We have also compared the accuracy of those expansions
with the approximations provided by the power series definition (6.2), the connection
formula (6.3) combined with (6.2) and also with Bühring’s formula [3, eq. 16.8.9] with
z0 = 1/2 as base point. The tables show the relative error provided by the different
approximations when we truncate the expansion after N terms.
The numerical results show that the expansions derived from a three-point Taylor
expansion (6.25) (with q = 0 or q = q1 ) are more accurate than the expansions derived
from a two-point Taylor expansion (6.18) or a standard Taylor expansion (6.5). But
moreover, the new expansions are more competititive than Bühring’s formula or the
standard Taylor series definition when z is close to 0 or when the difference between
any couple of the parameters (a, b1 , b2 , . . . , bp ) is close or equal to an integer. When z
is away from 0 the expansions derived from a three-point Taylor expansion (6.25) seems
to perform better than Bühring’s formula which, on the other hand, seems to perform
better than the expansion obtained using a one- or a two-point Taylor expansion. In any
case, expansion (6.25) with q = q1 seems to be the most accurate.
96 Chapter 6. A dual vision of uniform expansions
Approximation N =0 N =2 N =4 N =6
Power series definition (6.2) 8.971 · 10−2 2.829 · 10−3 1.368 · 10−4 7.742 · 10−6
Connection formula: (6.3)+(6.2) − − − −
Bühring’s formula z0 = 1/2 6.371 · 10−1 1.844 · 10−1 6.933 · 10−2 2.828 · 10−2
1-point expansion w = 1/2 (6.5) 4.702 · 10−2 5.770 · 10−4 7.870 · 10−4 1.127 · 10−7
2-point expansion q = 0 (6.18) 8.932 · 10−3 1.551 · 10−6 3.325 · 10−10 7.676 · 10−14
2-point expansion q = q0 (6.18) 7.280 · 10−4 3.829 · 10−8 2.201 · 10−12 2.376 · 10−16
3-point expansion q = 0 (6.25) 1.943 · 10−4 8.069 · 10−11 4.189 · 10−17 2.345 · 10−23
3-point expansion q = q1 (6.25) 2.155 · 10−6 4.879 · 10−13 1.190 · 10−19 3.003 · 10−26
Approximation N =0 N =2 N =4 N =6
Power series definition (6.2) − − − −
Connection formula: (6.3)+(6.2) 8.182 · 10−2 2.442 · 10−3 1.384 · 10−4 9.510 · 10−6
Bühring’s formula z0 = 1/2 1.101 · 10−1 1.361 · 10−3 2.099 · 10−5 3.474 · 10−7
1-point expansion w = 1/2 (6.5) 3.474 · 10−1 1.021 · 10−1 3.301 · 10−2 1.115 · 10−2
2-point expansion q = 0 (6.18) 2.705 · 10−1 5.812 · 10−2 1.645 · 10−2 5.089 · 10−3
2-point expansion q = q0 (6.18) 2.799 · 10−2 1.134 · 10−3 5.157 · 10−5 2.476 · 10−6
3-point expansion q = 0 (6.25) 4.963 · 10−2 6.413 · 10−4 1.053 · 10−5 1.873 · 10−7
3-point expansion q = q1 (6.25) 3.102 · 10−3 1.518 · 10−5 8.371 · 10−8 4.880 · 10−10
Table 6.2: p = 3, a = 1, ~b = 1 4 5 5 7 5
, ,
2 3 6
, ~c = , ,
3 5 7
, z = −3 + i.
Approximation N =0 N =2 N =4 N =6
Power series definition (6.2) 3.701 · 10−1 1.598 · 10−1 9.907 · 10−2 7.070 · 10−2
Connection formula: (6.3)+(6.2) 4.657 · 10−1 1.722 · 10−1 1.029 · 10−1 7.220 · 10−2
Bühring’s formula z0 = 1/2 6.007 · 10−1 1.757 · 10−1 6.512 · 10−2 2.606 · 10−2
1-point expansion w = 1/2 (6.5) 2.393 · 10−1 7.884 · 10−2 2.952 · 10−2 1.171 · 10−2
2-point expansion q = 0 (6.18) 1.959 · 10−1 1.416 · 10−2 1.222 · 10−3 1.125 · 10−4
2-point expansion q = q0 (6.18) 4.168 · 10−2 1.167 · 10−3 3.610 · 10−5 1.169 · 10−6
3-point expansion q = 0 (6.25) 2.256 · 10−2 1.085 · 10−4 6.332 · 10−7 3.959 · 10−9
3-point expansion q = q1 (6.25) 4.522 · 10−3 1.232 · 10−5 3.713 · 10−8 1.172 · 10−10
iπ
Table 6.3: p = 3, a = 1, ~b = 1 4 5 5 7 5
, ,
2 3 6
, ~c = , ,
3 5 7
, z=e4.
Approximation N =0 N =2 N =4 N =6
Power series definition (6.2) 1.179 · 10−1 5.519 · 10−3 4.757 · 10−4 5.429 · 10−5
Connection formula: (6.3)+(6.2) − − − −
Bühring’s formula z0 = 1/2 − − − −
1-point expansion w = 1/2 (6.5) 1.081 · 10−1 2.746 · 10−3 8.254 · 10−5 2.695 · 10−6
2-point expansion q = 0 (6.18) 2.492 · 10−2 2.843 · 10−5 4.064 · 10−8 6.312 · 10−11
2-point expansion q = q0 (6.18) 3.416 · 10−3 1.276 · 10−6 5.058 · 10−10 2.032 · 10−13
3-point expansion q = 0 (6.25) 1.823 · 10−3 1.454 · 10−8 1.322 · 10−13 3.370 · 10−15
3-point expansion q = q1 (6.25) 6.574 · 10−4 2.401 · 10−9 7.157 · 10−15 3.265 · 10−15
7iπ
Table 6.4: p = 3, a = 1, ~b = (2, 3, 1), ~c = (2, 3, 4), z = 12 e 6 .
6.3 Final remarks and numerical experiments 97
Approximation N =0 N =2 N =4 N =6
Power series definition (6.2) 3.748 · 10−2 2.369 · 10−3 3.921 · 10−4 8.987 · 10−5
Connection formula: (6.3)+(6.2) − − − −
Bühring’s formula z0 = 1/2 4.515 · 10−2 2.352 · 10−3 2.963 · 10−4 5.377 · 10−5
1-point expansion w = 1/2 (6.5) 2.464 · 10−1 1.715 · 10−2 1.236 · 10−3 9.052 · 10−5
2-point expansion q = 0 (6.18) 1.200 · 10−2 3.011 · 10−5 1.244 · 10−7 5.885 · 10−10
2-point expansion q = q0 (6.18) 2.640 · 10−2 3.445 · 10−5 4.698 · 10−8 6.538 · 10−11
3-point expansion q = 0 (6.25) 1.797 · 10−3 6.212 · 10−8 3.113 · 10−12 2.047 · 10−16
3-point expansion q = q1 (6.25) 2.923 · 10−3 7.305 · 10−8 1.908 · 10−12 1.419 · 10−16
2iπ
Table 6.5: p = 7, a = 1, ~b = 1 1 3 5 3 7 5 12 2 5 6 6 9 5
, z = 23 e
, , , , , ,
2 3 4 7 8 6 9
, ~c = , , , , , ,
11 3 4 7 5 10 12
3 .
Approximation N =0 N =2 N =4 N =6
Power series definition (6.2) − − − −
Connection formula: (6.3)+(6.2) 9.630 · 10−3 1.115 · 10−4 4.577 · 10−6 2.786 · 10−7
Bühring’s formula z0 = 1/2 2.122 · 10−2 1.591 · 10−4 2.206 · 10−6 3.710 · 10−8
1-point expansion w = 1/2 (6.5) 5.809 · 10−1 2.132 · 10−1 8.042 · 10−2 3.075 · 10−2
2-point expansion q = 0 (6.18) 6.942 · 10−2 7.813 · 10−3 1.608 · 10−3 3.893 · 10−4
2-point expansion q = q0 (6.18) 1.779 · 10−1 8.749 · 10−3 4.497 · 10−4 2.359 · 10−5
3-point expansion q = 0 (6.25) 2.684 · 10−2 2.381 · 10−4 3.198 · 10−6 4.812 · 10−8
3-point expansion q = q1 (6.25) 1.057 · 10−1 1.809 · 10−3 3.248 · 10−5 5.975 · 10−7
Table 6.6: p = 7, a = 1, ~b = 1 1 3 5 3 7 5 12 2 5 6 6 9 5
, , , , , ,
2 3 4 7 8 6 9
, ~c = , , , , , ,
11 3 4 7 5 10 12
, z = −2 − 2i.
Chapter 7
A Convergent Laplace’s Method
for Integrals
with large |z| (for fixed arg z ∈ (−π/2, π/2)). Imagine that we want to obtain an ap-
proximation of F (z). On the one hand, we could apply the lemma of Watson introduced
in subsection 2.2.1 of chapter 2. On the other hand, we could obtain a uniformly con-
vergent expansion of F (z) by applying the theory of chapter 3 in the situation where
h(t, z) = e−zt and the interval of integration is unbounded:
with Z ∞
e−zt gn (t)dt = O z −n−1 ,
Rn (z) := as z → ∞.
0
This last order estimate means that the expansion (7.3) is an asymptotic expansion
of F (z) for large |z|. However, as we know from chapter 2, the expansion (7.3) is
usually divergent. This can be regarded as a penalty for interchanging summation
99
100 Chapter 7. A convergent Laplace’s method for integrals
and integration when the integration interval [0, ∞) is not contained in the disk of
convergence of the Taylor expansion of g(t).1
For example, we may consider the function F (z) = ez E1 (z), where E1 (z) is the
exponential integral and derive the expansion (2.15) which is asymptotic for large
|z|, but it is not convergent for any value of z ∈ C.
Typically, the function g(− log τ ) is analytic at τ = 1 and has a Taylor expansion
of the form
n−1
X
k (−1)k dk
g(− log τ ) = Ak (1 − τ ) + rn (τ ), Ak := [g(− log τ )]τ =1 ,
k=0
k! dτ k
that, due to the singularity at τ = 0, is convergent in any open disk Dr (1) centered
at τ = 1 with radius r ≤ 1. In the best scenario when r = 1 we can apply theorem
3.3.1, case 3. In this situation, the coefficients Ak can also be computed from the
Taylor coefficients of g(t) at t = 0 using formula (A.5) in corollary A.0.2 of appendix
A. Using theorem 3.3.1 we can interchange summation and integration and we find
Z ∞ Z 1 n−1
−zt z−1
X k!
F (z) := g(t)e dt = g(− log τ )τ dτ = Ak + Rn (z), (7.5)
0 0 k=0
(z)k+1
where (z)n denotes the Pochhammer’s symbol and the remainder Rn (z) is given by
Z 1
Rn (z) = τ z−1 rn (τ )dτ.
0
From (3.14), the remainder can be bounded by means of |Rn (z)| ≤ M Γ(n+1)Γ(z)
Γ(z+n+1)
with M > 0 independent of n and z. Therefore, we have
expansion obtained by applying the theory of chapter 3 is both convergent (and then,
the more terms considered the better) and asymptotic for large |z|. The main drawback
is that the coefficients of the expansion are more difficult to compute as they are the
Taylor coefficients of a composite function.
Expansions of the form (7.3) are named factorial series [145, §17.3] and they were
first studied by Newton and Stirling. Their relation with a compact Mellin transform
R 1 z−1
0
u f (u)du was found by Schlömilch [133] although it was Nörlund [99] who showed
the neccesary and sufficient conditions for a factorial series to be expressible as such
an integral. Moreover, Nielsen [97] showed the relation between factorial series and
asymptotic series by using the Stirling numbers [15, §26.8]. A comprehensive study
of factorial series was made in [34, 96]. However, in those manuscripts, the starting
point of the analysis is the factorial series itself. Therefore, neither formulas for the
coefficients of the approximation (7.5) nor error bounds are given. But moreover, the
integrals considered are of any of the forms given in (7.5) where the phase function (in
the unbounded integral representation of F (z)) is simpy f (t) = t. Generalizations of
asymptotic and convergent series to integrals of the form
Z ∞
F (z) := e−zf (t) g(t)dt, (7.6)
0
for a more general function f (t), are not considered in the literature.
The aim of the present chapter of this thesis is to fill this gap, that is, to derive a
new convergent and asymptotic expansion to the integral (7.6) with general functions
f (t) and g(t). We establish the conditions that the functions f (t) and g(t) must satisfy
and we estimate the remainder term. Furthermore, we obtain an explicit and simple
formula for the coefficients of the new expansion, in contrast to the classical Laplace
method, where the formulas for the coefficients of the expansion involve combinatorial
objects whose complexity increases with the numbers of terms considered in the expansion
[60, 152, 153].
where 0 < ρ ≤ 1, m ∈ N, <s > 0 and <z > x0 > 0 is large. We also assume that the
function f (x) satisfies one of the following two hypotheses:
Either the function f (x) is analytic inside an open disk of center 0 and radius r > ρ,
Theorem 7.1.1. Consider the integral (7.7) with the above mentioned hypotheses on the
function f (x). Then, for n = 1, 2, 3 . . .
n−1
1 X f (k) (0)
k+s
F (z) = Bρm ,z + Rn (z), (7.8)
m k=0 k! m
where By (w, z) is the incomplete beta function [116, §8.17]. The remainder Rn (z) can be
bounded in the form
M n+<s
(ρ ) n Bρ m
m
, <z if r > ρ,
0
|Rn (z)| ≤ M (z)B (n + <s, σ) if r = ρ < 1, (7.9)
M (z)B (n + <s, <z + σ − 1) if r = ρ = 1,
where B(w, z) is the beta function, M is a certain positive constant independent of n and
z, and M (z) > 0 is independent of n. On the other hand, in the first line of (7.9) we
have ρ0 := r − > ρ, with small. Therefore, expansion (7.8) is convergent, with an
exponential rate of convergence if r > ρ and a power rate if r = ρ. More precisely, as
n → ∞,
O (n−1 (ρ/ρ0 )n ) if r > ρ, ρ < 1,
O n−<z ρ−n
if r > ρ = 1,
0
Rn (z) = −σ
(7.10)
O (n ) if r = ρ < 1,
O n−(<z+σ−1)
if r = ρ = 1.
Expansion (7.8) is also an asymptotic expansion of F (z) for large |z|. The terms of the
expansion and the remainder term satisfy, as |z| → ∞,
n+s n+s n+s
Bρm , z = O z− m , Rn (z) = O z − m , n = 1, 2, 3, . . . (7.11)
m
7.1 An expansion for compact Mellin transforms 103
Proof. The key point of the proof is an accurate analysis of the remainder rn (x) of the
Taylor series expansion of f (x) at x = 0, similar to the one carried out in section 3.2 of
chapter 3.
Consider the Taylor series expansion of f (x) at x = 0:
n−1 (k)
X f (0)
f (x) = xk + rn (x), |x| < r. (7.12)
k=0
k!
with Z ρ
Rn (x) := (1 − xm )z−1 xs−1 rn (x)dx. (7.14)
0
Then, (7.8) follows from (7.13) by performing the change of variables x 7→ y given
by xm = y and using the integral definition [116, eq. 8.17.1] of the incomplete beta
function.
On the other hand, the remainder rn (x) can be written in the form
xn
I
f (w)dw
rn (x) = , x ∈ [0, ρ), (7.15)
2πi C wn (w − x)
where C is a simple closed circle around the origin of radius ρ0 := r − > ρ
encircling the points w = 0 and w = x in the positive direction. Then
xn xn
I I
f (w) f (w)
|rn (x)| ≤ dw = dw, x ∈ [0, ρ). (7.16)
2π C wn (w − x) 2πρn0 C w − x
As the function f (x) is analytic in the disk D0 (r), its modulus is bounded in C by a
positive constant independent of x and n, and so is the factor 1/(w − x). Therefore,
xn
|rn (x)| ≤ M̄ ,
ρn0
Consider now the case r = ρ. The function f (x) may have an integrable singularity
at x = ρ. We proceed in the same way as we did in the previous point, considering
the Taylor series expansion of f (x) at x = 0 and interchanging summation and
integration in (7.7) to find the right hand side of (7.8) with Rn (z) defined in (7.14)
and rn (x) given by the Cauchy’s integral (7.15). However, we have to be careful
when bounding the remainder rn (x) as the function f (x) no longer is analytic in
104 Chapter 7. A convergent Laplace’s method for integrals
the closed disk D0 (ρ) and therefore the integral on the right hand side of (7.16)
no longer can be bounded by a positive constant. Nevertheless, we note that the
integral (7.15) is a constant function of that is defined for = 0 (that is, when
we set ρ0 = ρ) and it is a continuous function of since it is the integral of an
integrable function. Thus, formula (7.15) is also valid for r = ρ if we take the limit
→ 0 and we consider that C is a circle of radius r = ρ. Moreover, we can perform
a translation of the integration variable w 7→ v + ρ to find
xn
I
f (v + ρ)dv
rn (x) = , x ∈ [0, ρ),
2πi C̄ (v + ρ)n (v + ρ − x)
xn (ρ − x)σ−1 |u|σ−1
I
|rn (x)| ≤ M̄ du, (7.17)
ρn C̄/(ρ−x) |u + 1|
where the integration contour C̄/(ρ−x) is the scaled circle of center u = −ρ/(ρ−x)
and radius ρ/(ρ − x) traversed in the positive direction. In other words,
ρ ρ
C̄/(ρ − x) = u∈C: u+ = .
ρ−x ρ−x
In the limit x → ρ, the scaled circle becomes the imaginary axis traversed upwards
and the integral along this path in formula (7.17) is finite. Therefore, the right
hand side of (7.17) can be bounded in the form
xn (ρ − x)σ−1
|rn (x)| ≤ M̃ , x ∈ [0, ρ),
ρn
for a certain M̃ > 0 independent of n and x. Introducing this bound for the
remainder rn (x) into (7.14) we get
Z ρ
M̃
|Rn (z)| ≤ n (1 − xm )<z−1 xn+<s−1 (ρ − x)σ−1 dx. (7.18)
ρ 0
with M (z) := ρ<s+σ−1 M̃ M̄ (z), and we obtain the second line of (7.9).
– If ρ = 1 we find
Z 1
|Rn (z)| ≤ M̃ xn+<s−1 (1 − x)<z+σ−2 (1 + x + . . . + xm−1 )<z−1 dx.
0
When <z > 1 we have the bound (1 + x + . . . + xm−1 )<z−1 ≤ m<z−1 whereas
when <z < 1 we find (1 + x + . . . + xm−1 )<z−1 ≤ 1. Therefore, (1 + x + . . . +
xm−1 )<z−1 ≤ m̃(z) := max{1, m<z−1 }, independent of n, and then
Z 1
|Rn (z)| ≤ M̃ m̃(z) xn+<s−1 (1−x)<z+σ−2 dx = M (z)B (n + <s, <z + σ − 1) ,
0
In the first integral on the right hand side of (7.19) we use that (ρ−x)σ−1 ≤ [ρ(1 − ε)]σ−1 ,
for all x ∈ [0, ερ]. In the second integral we use that, for <z > 1, (1 − xm )<z−1 ≤
(1 − (ερ)m )<z−1 for all x ∈ [ερ, ρ]. Then
Z ρ Z ερ
m <z−1 n+<s−1 σ−1
(1 − x ) x σ−1
(ρ − x) dx ≤ [ρ(1 − ε)] (1 − xm )<z−1 xn+<s−1 dx +
0 0
Z ρ
m <z−1
+ [1 − (ερ) ] xn+<s−1 (ρ − x)σ−1 dx.
ερ
The last integral above can be bounded by a positive constant, say K, independent
of z, and the first integral is an incomplete beta function. Therefore
Z ρ
(1 − xm )<z−1 xn+<s−1 (ρ − x)σ−1 dx ≤
0
[ρ(1 − ε)]σ−1
n + <s
≤ B(ερ)m , <z + K [1 − (ερ)m ]<z−1 .
m m
For large |z|, the second term on the right hand side above is exponentially small com-
−(n+s)/m
pared to the first one, that is of the order O z . Therefore, the asymptotic
behavior (7.11) for the remainder Rn (z) also follows in the case r = ρ and expansion
(7.8) is not only convergent, but asymptotic, as |z| → ∞.
Example 2. Consider the integral representation of the Bessel Jν (z) function [107, eq.
10.9.4], given by
Z 1
2(z/2)ν
Jν (z) = 1/2 (1 − x2 )ν−1/2 cos(zx)dx,
π Γ(ν + 1/2) 0
for large ν. To obtain a convergent and asymptotic expansion for large ν of the function
Jν (z) we can apply theorem 7.1.1 with m = 2 and f (x) := cos(zx). The function f (x) is
entire and then r > ρ := 1. We also have f (2k+1) (0) = 0 and f (2k) (0) = (−1)k z 2k , for all
k = 0, 1, 2, . . .. Therefore, using approximation (7.8) we find
" ∞ #
2(z/2)ν 1 X (−1)k z 2k
2k + 1 1
Jν (z) = 1/2 B ,ν +
π Γ(ν + 1/2) 2 k=0 (2k)! 2 2
∞
(z/2)ν X (−1)k z 2k Γ(k + 12 )
=
π 1/2 k=0 Γ(2k + 1) Γ(k + ν + 1)
Using the duplication formula [4, eq. 5.5.5] for the gamma function Γ(2k + 1) and after
some simplifications we find the power series definition of Jν (z) [107, eq. 10.2.2], which
is a convergent series that is also asymptotic for large |ν|.
point t = 0 of the integration interval. Moreover, if b is finite we also allow a branch point
at the other end point t = b whereas if the integration interval is unbounded we let an
exponential growth of the integrand at infinity. More precisely, we consider generalized
Laplace’s transforms of the form
Z b
m
F (z) := e−zt ts−1 (1 − t/b)σ−1 h(t)dt, (7.20)
0
with m ∈ N, <s > 0, 0 < <σ ≤ 1, 0 < b ≤ ∞ and large |z| with | arg z| < π/2. That is,
we assume that <z ≥ x0 for some fixed x0 > 0. Moreover, in the case b = ∞ we assume
that the factor (1 − t/b)σ−1 is replaced by 1 and that h(t) satisfies some growth condition
at ∞ to ensure the convergence of the integral (7.20).
The main idea to obtain a convergent and asymptotic approximation of the integral
transform F (z) given in (7.20) is to reduce the study to one of the cases analyzed in
theorem 7.1.1 by performing a logarithmic change of variables. More precisely, we take
1/m
log(1 − xm )
t=x − , (7.21)
xm
where t > 0 for x > 0. The inverse of the transformation is given by
m 1/m
1 − e−t
x=t , (7.22)
tm
with x > 0 for t > 0. As we will see below, this change of variables transforms the
integral (7.20) to one of the form considered in theorem 7.1.1. Moreover, in lemma A.0.1
of appendix A we can find a formula for the composite function after the change of
variables (7.21).
Before deriving an expansion for the integral F (z) on the right hand side of (7.20),
we must give some insight in the transformation (7.21)-(7.22). To this end, we define the
following region:
Definition. For any 0 < b ≤ ∞ and m ∈ N, we define the open complex region
m m
Sm (b, 0) := t ∈ C : 1 − e−t < ρ , ρ := 1 − e−b ≤ 1,
(7.23)
where only the branch that contains the real interval (0, b) is considered (see figure 7.1).
(The “extra” argument 0 in the notation of Sm (b, 0) will be clear later). Observe that
ρ < 1 when 0 < b < ∞ and ρ = 1 if and only if b = ∞. Moreover, for any b1 < b2 we
have that Sm (b1 , 0) ⊂ Sm (b2 , 0) ⊂ Sm (∞, 0).
The importance of the region Sm (b, 0) in the following analysis is that its image under
the transformation (7.21) is a disk of center x = 0 and radius ρ1/m .
Now, we obtain a convergent and asymptotic expansion for integrals of the form
(7.20). To this end, we assume one of the following hypotheses for the function h(t):
Either b < +∞ and the function h(t) is analytic in Sm (b0 , 0) for some b0 > b,
Figure 7.1: Region Sm (b, 0) for m = 2 (left), m = 3 (middle) and m = 4 (right) with b =
∞ (larger blue unbounded regions, including green areas) and finite b (inner green bounded
figures). For any value of b, finite or infinite, the boundary of these figures is comprised by
the curve t(θ) parametrized in the form t(θ) = [− log(1 − ρeimθ )]1/m , −π < θ ≤ π, selecting
the continuous branch for which t > 0 for θ = 0; that is, t(θ) = | log(1 − ρeimθ )|1/m eiθ ,
−π < θ ≤ π. In these figures we have highlighted four points of the boundary of Sm (b, 0):
A := t(0) = [− log(1 − ρ)]1/m > 0, B := t(π/2) = [− log(1 − ρeimπ/2 )]1/m = i|B|, C :=
t(π/m) = [− log(1 + ρ)]1/m = eiπ/m |C|, P := t(2π/m) = e2iπ/m A. In this figures we have
depicted these points for ρ < 1 (green region with b < ∞). When b → ∞ (ρ → 1) we have that
A → +∞, P → e2iπ/m ∞ and C → (log 2)1/m eiπ/m ∀ m. Roughly speaking, the region Sm (b, 0)
is a circle around the origin of radius |C| = | log(1 + ρ)|1/m spiked along the m rays determined
by the m−th roots of the unity e2πik/m , k = 0, 1, . . . , m − 1 (the point A is on the first ray, the
point P is on the second one, the point C is on the middle angle between A and P ). For b = ∞
the (blue + green) regions are unbounded, as those spikes go up to the infinity. For b < ∞
the (green) regions are similar, but bounded, as those spikes are bounded, and contained in
Sm (∞, 0): observe that Sm (b1 , 0) ⊂ Sm (b2 , 0) for b1 < b2 .
(α)
where Bn (x) are the generalized Bernoulli polynomials of order α [36, §24.16], [88, Ch.
VI], and the factor (1 − t/b)σ−1 is replaced by 1 if b = ∞.
7.2 A convergent version of a generalized Watson’s lemma 109
for a certain ρ0 > ρ > 0, M > 0 independent of z and n and M (z) > 0 independent of
n. Therefore, expansion (7.24) is convergent, with an exponential rate of convergence for
b < +∞ and σ = 1, and a power rate otherwise. More precisely, as n → ∞,
−1 n
O (n (ρ/ρ0 ) ) if b < +∞ and σ = 1,
Rn (z) = O (n−σ ) if b < +∞ and σ 6= 1, (7.28)
−(<z−α)
O n if b = +∞.
Expansion (7.24) is also an asymptotic expansion of F (z) for large |z|: The terms of the
expansion (7.24) and the remainder satisfy
k+s k+s n+s
Bρ , z = O z− m Rn (z) = O z − m , as z → ∞ (7.29)
m
Proof. We consider the change of variables t 7→ x given in (7.21). Under this transfor-
mation, the region Sm (b, 0) for the variable t is transformed into the open disk of center
m
x = 0 and radius ρ1/m for the variable x, where ρ := 1 − e−b ≤ 1. The integral (7.20) is
transformed into
Z ρ1/m
F (z) = (1 − xm )z−1 xs−1 h̃(x)dx, (7.30)
0
with h̃(x) defined in (7.25). The analyticity hypotheses on the function h(t) guarantee
that the function h̃(x) satisfies one of the two analyticity conditions for the function f (x)
in theorem 7.1.1. Then, we can apply theorem 7.1.1 to this integral with f (x) replaced
by h̃(x) to find the approximation (7.24). The formula for the Taylor coefficients Ak
σ−1
follows from lemma A.0.1 with λ = ms − 1 and φ(t) = 1 − bt h(t) and the bounds for
the remainder are a consequence of (7.9) in theorem 7.1.1. More precisely we find:
When b < +∞ and σ = 1, we have ρ < 1 and the function h̃(x) is analytic in a
1/m m
certain open disk D0 (r) of radius r > ρ0 > ρ1/m , where ρ0 := 1 − e−b0 and the
first line of (7.27) follows from the first line of (7.9).
If b < +∞ and σ 6= 1 we have that ρ < 1 and the function h̃(x) has a branch point
at ρ1/m < 1. Therefore, the function h̃(x) is analytic in the open disk D0 (ρ1/m ) and
the third line of (7.9) implies the second line of the bound (7.27).
In the case b = +∞ we have ρ = 1 and the middle factor of h̃(x) in (7.25) is replaced
by 1. The function h̃(x) is analytic in the open unit disk D0 (1) and satisfies the
growth condition h̃(x) = O ((x − 1)−α ) as x → 1. Therefore, the fourth line of
(7.9) implies the third line of the bound (7.27) with σ replaced by 1 − α.
110 Chapter 7. A convergent Laplace’s method for integrals
To illustrate the applicability of theorem 7.2.1 we consider a function that will be use-
ful in the next sections to derive a convergent and asymptotic expansion of the parabolic
cylinder function [144].
− 1 " 1 #a− 12
log(1 − x2 ) 2 log(1 − x2 ) 2
h̃(x) = − 1−x −
x2 x2
k+1 k+1
j B ( 2 ) (1) j B ( 2 ) (1)
b k2 c b k2 c
k−2j
X (−1) j d a−1/2
X (−1) j 1
Ak = (1 − t) t=0
= −a .
j=0
j! (k − 2j)! dtk−2j j=0
j! (k − 2j)! 2 k−2j
Table 7.1 contains a numerical experiment about approximation (7.31) that shows its
convergent and asymptotic character for large |z|.
Table 7.1: Relative error [143, eq. 3.1.9] provided by the right hand side of (7.31) when
we truncate the series after n terms, for different values of z and a = 1.8.
7.3 Some magic tricks 111
2ε
2ε 2ε
Re(z) Re(z) Re(z) Re(z)
2ε 2ε 2ε 2ε
2ε 2ε
Figure 7.2: Region sm (b, 0) for m = 2 (left), m = 3 (middle left), m = 4 (middle right) and
m = 6 (right). In every picture, the m rays in the directions e2ikπ/m , k = 0, 1, 2, ..., m − 1, have
√
width 2ε and length bε/ m log 2 .
that we will approximate later and is related with the parabolic cylinder function [144].
The function h(t) = (t + 1)a−1/2 possesses a branch point at t = −1 that belongs to the
region S2 (∞, 0) (see figure 7.1 left) and then theorem 7.2.1 cannot be applied. However,
as we will see below, after some manipulations, the analyticity requirement in theorem
7.2.1 for the function h(t) can be relaxed, requiring its analyticity in a smaller region and
enlarging in this way the range of applicability of theorem 7.2.1.
In order to understand what manipulation we need, we first give some insight into the
geometry of the transformation t 7→ x defined in (7.21) with inverse given in (7.22). The
key point is that, after the change of variables t 7→ x, the original integration interval
[0, b) is transformed into a new integration interval in (7.30): x ∈ [0, ρ1/m ) ⊂ [0, 1) for
m
all b > 0, where ρ := 1 − e−b . Expansion (7.24) follows after a Taylor expansion of the
function h̃(x) given in (7.25) at x = 0 and an interchange of summation and integration
in (7.30). Then, the resulting expansion (7.24) is convergent whenever h̃(x) is analytic
in the disk D0 (ρ1/m ), that contains the integration interval. But h̃(x) is analytic in that
disk whenever h(t) is analytic in the image of the disk D0 (ρ1/m ) under the transformation
(7.22), that is the region Sm (b, 0).
Then, the trick to relax the analyticity condition of theorem 7.2.1 is a dilatation
of the original integration variable t ∈ [0, b) that squeezes the region Sm (b, 0) into a
smaller region sm (b, 0). (Again, the “extra” argument 0 in the notation will be clear in a
moment.) Such a region sm (b, 0) may be a starlike domain √ with center at t = 0 and with
m
m arbitrarily narrow spikes of width 2ε and length bε/ log 2 in the directions defined
by the m-th roots of the unity, with ε > 0 arbitrarily small (see Figure 7.2).
112 Chapter 7. A convergent Laplace’s method for integrals
Im(t)
Im(t)
Im(t) C2 C1
C1
C1
Re(t)
C2
Re(t)
Re(t)
C2
C3 C3 C4
Figure 7.3: The thick red star-like regions Λsm (∞, 0) contain the corresponding regions
Sm (∞, 0) limited by the blue curves; for m = 2 (left), m = 3 (middle) and m = 4
(right). In these figures C1 , C2 ,..., Cm are the m different m−th roots of − log(2), that is,
p p
Ck = |C|ei(2k+1)π/m , k = 0, 1, 2, ..., m − 1, |C| := m log(2) and Λ = m log(2) /ε for the given
ε > 0 used in the definition (7.33) of sm (b, 0).
2iπk/m bε
sm (b, 0) := t ∈ C : t = (r + iy)e ;0 ≤ r < √ ; k = 0, 1, . . . , m − 1; |y| < ε .
| m log 2|
(7.33)
Now, consider a certain dilatation parameter Λ > 1. After the dilatation u 7→ t given
by t = Λu, the region sm (b, 0) for the variable u becomes the larger region Λsm (b, 0) for
the new variable t. And the region Λsm (b, 0) contains the region Sm (b, 0) for any b > 0
whenever εΛ is larger than the distance from the origin t = 0 to the complementary of
the region Sm (∞, 0). That distance is attained at the intersecion of the boundary of
Sm (∞, 0) with √the rays t = rei(2k+1)π/m , r > 0, k = 0, 1, . . . , m − 1, which occurs at a
distance C = m log 2 (see figure 7.3).
Therefore,
√ the region Λsm (b, 0) contains the region Sm (b, 0) for any b > 0 whenever
Λ > m log 2 /ε. Then
Z b Z Λb σ−1
−zum s−1
u σ−1 1 −(z/Λm )tm s−1 t
F (z) := e u 1− h(u)du = s e t 1− h̄(t)dt,
0 b Λ 0 Λb
with h̄(t) := h(t/Λ). Whenever h(u) is analytic in sm (b, 0), h̄(t) is analytic in Sm (b, 0) ⊂
Λsm (b, 0) and theorem 7.2.1 may be applied to this last integral with z replaced by z/Λm
and h replaced by h̄.
The key point of why the dilatation trick discussed above to enlarge the applicability of
theorem 7.2.1 works is hidden in the shape of the region Sm (∞, 0): any ray arg t = θ will
eventually cut the boundary of the region Sm (∞, 0) unless θ = 2kπ/m, k = 0, 1, . . . , m−1.
Then, for sufficiently large Λ, the dilatation u 7→ t = Λu will move the singularities of
h(t) away from the region Sm (b, 0) except for those singularities located in the axes of
the starlike region sm (b, 0) as they are invariant under the dilatation no matter how
large Λ is. Nevertheless, in the case b = ∞ we can still use another trick to avoid those
singularities: an appropriate rotation of the original integration path [0, ∞) whenever
the function h(t) is analytic in that sector. The effect of this rotation is a rotation of
the axes so that we can avoid the singularities located there. Then, choose an angle θ
m
such that | arg z + mθ| < π/2 (ideally θ = − arg m
z
if arg z 6= 0). If e−αt h(t) is bounded
7.3 Some magic tricks 113
as t → ∞ in the sector arg t ∈ [0, θ] (and not only for t > 0 as it was required in the
hypothesis of theorem 7.2.1) and h(t) is analytic in that sector, we can invoke Cauchy’s
theorem to rotate the path of integration that angle θ. Then
Z ∞
imθ m
F (z) = eisθ
e−ze t ts−1 h(eiθ t)dt.
0
Now, the singularities of h(eiθ t) are not located on the rays arg t = 2kπ/m, k =
0, 1, . . . , m − 1. Therefore, we can (if necessary) perform a dilatation Λ > 1 and ap-
ply theorem 7.2.1 afterwards.
We can combine both tricks to enlarge the range of applicability of theorem 7.2.1.
The result is summarized in the following theorem.
Theorem 7.3.1. Consider the integral
Z b
m
F (z) := e−zt ts−1 (1 − t/b)σ−1 h(t)dt, (7.34)
0
with m ∈ N, <s > 0, 0 < <σ ≤ 1, 0 < b ≤ +∞ and large |z|, that is, <z ≥ x0 > 0,
assuming that the factor (1 − t/b)σ−1 is replaced by 1 if b = +∞. Assume also one of the
following hypotheses for the function h(t):
Whether the function h(t) is analytic in the starlike region sm (b, 0) defined in (7.33)
m
and, if b = +∞, the function e−αt h(t) is bounded as t → +∞, for a certain
α < min{1, x0 }.
m
Or b = +∞ and there exists an angle θ with | arg z +mθ| < π/2 such that e−αt h(t)
is bounded as t → ∞ in the sector arg t ∈ [0, θ] for a certain α < min{1, x0 }, and
the function h(t) is analytic in the sector arg t ∈ [0, θ] and also in the starlike region
sm (∞, 0) defined in (7.33) rotated an angle θ:
sm (∞, θ) := t ∈ C : t = (r + iy)ei(θ+2πk/m) ; r ≥ 0; k = 0, 1, . . . , m − 1; |y| < ε,
(7.35)
for a certain ε > 0.
Then, for n = 1, 2, 3, . . .
n−1
1 X k+s z
F (z) = Ak Bρ(Λ) , + Rn (z), (7.36)
mΛs k=0 m Λm
√
for any Λ ∈ C with |Λ| > m log 2 /ε and arg(Λ) = θ, with θ = 0 if h(t) is analytic in
m
sm (b, 0). The parameter ρ(Λ) is defined by ρ(Λ) = 1 − e−(|Λ|b) and Ak are the Taylor
coefficients of
s −1 1 !σ−1 1!
log(1 − xm ) m log(1 − xm ) m log(1 − xm ) m
x x
h̃(x) := − 1− − h −
xm bΛ xm Λ xm
at x = 0, with the middle factor replaced by 1 when b = +∞. These Taylor coefficients
Ak can be computed either, directly using an algebraic manipulator, or by means of the
following formula (see lemma A.0.1):
k
bm c
j ( k+s
m ) k−mj
" σ−1 #
X (−1) B j (1) d t
Ak = k−mj k−mj
1− h(t) , (7.37)
j=0
j! (k − mj)! Λ dt b
t=0
114 Chapter 7. A convergent Laplace’s method for integrals
(α)
where the factor (1 − t/b)σ−1 is replaced by 1 if b = +∞ and Bk (x) are the generalized
Bernoulli polynomials of order α.
The remainder term Rn (z) can be bounded in the form
M n+<s z
ρ n Bρ(Λ) m
, < Λ m if b < +∞ and σ = 1,
0
|Rn (z)| ≤ M (z)B (n + <s, <σ) if b < +∞ and σ 6= 1, (7.38)
M (z)B n + <s, < z − α
Λm
if b = +∞,
for a certain ρ0 > ρ > 0, M > 0 independent of z and n and M (z) > 0 independent of
n. Therefore, expansion (7.36) is convergent, with an exponential rate of convergence for
b < +∞ and σ = 1, and a power rate otherwise. More precisely, as n → ∞,
O (n−1 (ρ/ρ0 )n ) if b < +∞ and σ = 1,
−σ
Rn (z) = O (n ) if b < +∞ and σ 6= 1, (7.39)
O n−(<( Λzm )−α)
if b = +∞.
Example 4. Consider the function Ū2 (a, z) defined in (7.32) and given by
Z ∞
2
Ū2 (a, z) := (t + 1)a−1/2 e−zt dt, <a > −1/2, <z > 0.
0
dn h i
1
where we have used that h(t) = (−1)n 2
− a n . The first few coefficients are
dtn t=0
The main idea to obtain a convergent and asymptotic Laplace’s method for integrals
is to use the idea of the modified Laplace’s method introduced in [74] and summarized in
section 2.2.3 of chapter 2, that is, to split the phase function f (t) into its asymptotically
dominant monomial (as |z| → ∞) and a subdominant remainder. Then, we obtain
a monomial in the exponent of the exponential and theorems 7.2.1 or 7.3.1 may be
applied. We clearly state the hypotheses H7.4.(i)-H7.4.(vi) that the functions f (t) and
g(t) must satisfy in order for the new Laplace’s method for integrals to be convergent
and asymptotic, as |z| → ∞.
116 Chapter 7. A convergent Laplace’s method for integrals
m f (m) (t0 )
p(t) := f (t0 ) + η(t − t0 ) , η := > 0.
m!
Roughly speaking, p(t) is the asymptotically dominant monomial of the phase
function f (t), as |z| → ∞ whereas the remainder term fm (t) := f (t) − p(t) is
subdominant:
∞
X f (k) (t0 )
fm (t) := f (t) − p(t) = (t − t0 )k , |t − t0 | < r,
k=p
k!
Remark 7.4.1. Hypotheses H7.4.(i), H7.4.(ii) and H7.4.(vi) are similar to the hypotheses
required for the functions f (t) and g(t) in order to apply the classical method of Laplace
for integrals (see subsection 2.2.2) whereas hypotheses H7.4.(iii), H7.4.(iv) and H7.4.(v)
will guarantee the convergence of the expansion.
Then, according to hypothesis H7.4.(ii) the integral (7.42) can be written in the form
Z ∞
−zf (t0 ) m
F (z) = e e−zη(t−t0 ) h(t, z)ta−1 dt, h(t, z) := e−zfm (t) g(t), (7.43)
0
where the exponent of the exponential function consists only of the asymptotically rele-
vant monomial of the phase function, as |z| → ∞. On the other hand, the subdominant
part fm (t) is included, together with g(t), in the function h(t, z). We split the integral
(7.43) at t = t0 to find
Z t0 Z ∞
−zf (t0 ) −zη(t−t0 )m −zf (t0 ) m
F (z) = e e a−1
h(t, z)t dt + e e−zη(t−t0 ) h(t, z)ta−1 dt.
0 t0
7.4 A convergent and asymptotic Laplace’s method 117
If t0 = 0 the first integral vanishes whereas if t0 > 0, m is even and we can perform a
shift in the integration variable of the form t0 − t 7→ t. On the other hand, no matter the
value of t0 , in the second integral above we perform a shift of the form t − t0 7→ t. We
get
Z t0 Z ∞
−zf (t0 ) −zηtm −zf (t0 ) m
F (z) = e e h(t0 −t, z)(t0 −t) a−1
dt + e e−zηt h(t+t0 , z)(t+t0 )a−1 dt.
0 0
With the aim of treating both integrals above at the same time, we define b− := t0 and
b+ = +∞ and we define the functions
Z b±
m
±
F (z) := e−zηt h (t0 ± t, z) (t0 ± t)a−1 dt. (7.44)
0
Then
F (z) = e−zf (t0 ) F − (z) + F + (z) .
(7.45)
Now, hypothesis H7.4.(iii) assures that we can apply theorem 7.2.1 to the function F − (z)
whereas hypotheses H7.4.(iv) and H7.4.(v) guarantee that we can perform a dilatation
and a rotation and apply theorem 7.3.1 to F + (z). But, the function h considered in the-
orems 7.2.1 and 7.3.1 does not depend on the asymptotic variable z and now the function
h(t, z) in the integrand of both integrals F ± (z) does. Therefore, we are introducing a
new ingredient in the analysis: the function h and therefore its Taylor coefficients Ak do
depend on the asymptotic variable z.
On the one hand, this dependence on the variable z does not have any influence
on the convergence of expansion (7.24) or (7.36) of theorems 7.2.1 and 7.3.1 and the
corresponding convergence rate for the remainder Rn (z) of those expansions (7.28) and
(7.39), as n → ∞, remain valid. On the other hand, that dependence on z does have an
effect on the asymptotic behavior of those expansions and formulas (7.11) are no longer
valid. That is, when the function h depends on z the asymptotic features as |z| → ∞ of
the new expansion must be proved.
The situation is similar to the one analyzed in [74] as the function h(t, z) also depends
on the asymptotic variable z. This dependence is translated to the Taylor coefficients.
However, the asymptotic behavior, as |z| → ∞, of the coefficients Ak is known and it
does not disrupt the asymptotic character of the expansion.
We have the following result, based on the fact that the n−coefficient of the Taylor
expansion at t = t0 of the function h(t, z) defined in (7.43) is a polynomial in z of degree
bn/pc, where p > m denotes the first non-vanishing derivative of f (t) at t = t0 after the
m−derivative (see subsection 2.2.3).
Lemma 7.4.2. Let f (t) and g(t) be analytic functions at t = t0 and define the function
h(t, z) := e−zfm (t) g(t) given in (7.43) with fm (t) defined in hypothesis H7.4.(ii). For any
m ∈ N and Λ, λ, µ ∈ C, consider the function
λ −1 1 !µ−1
log(1 − xm ) m log(1 − xm ) m
x
h̃(x, z) := − t0 + − ×
xm Λ xm
1 ! (7.46)
log(1 − xm ) m
x
× h t0 + − ,z .
Λ xm
118 Chapter 7. A convergent Laplace’s method for integrals
Proof. The Taylor coefficients An (z) of h̃(x, z) can be computed using formula (A.2) in
lemma A.0.1. Then, in the computation of the n−th coefficient An (z), the variable z
only appears in the derivatives of the function h(t, z). But it has been shown in [74] that
the k−th Taylor coefficient of h(t, z) at t = t0 is in fact a polynomial in z of degree b kp c.
As the range of the index of summation in (A.2) goes from k = 0 to k = b np c, we conclude
that the Taylor coefficients An (z) of h̃(x, z) at x = 0 are also polynomials in z of degree
b np c.
Finally, we give the main result of this chapter of the thesis: a convergent and asymp-
totic expansion of the integral (7.42) with an explicit formula for the coefficients of the
expansion and with estimates for the remainder. The main idea to derive it is to write the
integral (7.42) in the form (7.44)-(7.45) and apply theorem 7.2.1 to F − (z) and theorem
7.3.1 to F + (z). However, as the function h(t, z) depends on the asymptotic variable z
we must prove that the new expansion is still asymptotic, as |z| → ∞. The proof of this
property is technical and somewhat long and tedious. We have:
Theorem 7.4.3. Consider the integral (7.42) and assume hypotheses H7.4.(i)-H7.4.(vi)
given above with the function h(t, z) defined in (7.43) and the parameters m, t0 and η
given in hypothesis H7.4.(ii). Then, for n = 1, 2, 3, . . .
n
bmc n+λ±
( )
X (−1)k B m
(1) (±1)n−km dn−km h µ± −1 i
A±
n (z) =
k
t h(t, z) , (7.49)
k=0
(Λ± )n−km k! (n − km)! dtn−km t=t0
(α)
where Bn (x) are the generalized Bernoulli polynomials [36, §24.16], [88, Ch. VI].
7.4 A convergent and asymptotic Laplace’s method 119
The remainder Rn (z) is bounded by the sum of the second line of (7.27) with s = 1
and the third line of (7.27) with s = λ+ . Therefore, expansion (7.47) is convergent. It
is also an asymptotic expansion of F (z) for large |z|: On the one hand,
the terms of the
k k+1 k k+λ+
expansion between brackets inside the sum (7.47) are of order O z p − m + z p − m ,
n n+1 n n+λ+
− −
as |z| → ∞. On the other hand, the remainder Rn (z) = O z p m + z p m , as
|z| → ∞, with p given in H7.4.(ii).
Proof. We split the phase function f (t) in the form given in hypothesis H7.4.(ii), and
we write the integral (7.42) in the form (7.44)-(7.45). Using hypothesis H7.4.(iii) we can
apply theorem 7.2.1 to F − (z) with b = t0 , s = 1, σ = a − d<ae + 1 and h(t) replaced by
ta−1
0 h(t0 − t, z)(1 − t/t0 )
a−σ
= ta−1
0 h(t0 − t, z)(1 − t/t0 )
d<ae−1
and z replaced by ηz. On
the other hand, using hypotheses H7.4.(iv) and H7.4.(v) we can apply theorem 7.3.1 to
F + (z) with b = +∞, σ = 1, z replaced by ηz and either s = 1 with h(t) replaced by
h(t0 + t, z)(t0 + t)a−1 if t0 > 0, or s = a and h(t) replaced by h(t0 + t, z) if t0 = 0. In
this way, we find the expansion (7.47) being A± ±
k (z) the Taylor coefficients of h̃ (x, z) at
x = 0. Formula (7.49) for the Taylor coefficients A± n (z) follows from lemma A.0.1 with
λ±
±
t µ −1 t
λ = m − 1 and φ(t) = t0 ± Λ± h(t0 ± Λ± ).
The remainder Rn (z) is given by the sum of the remainders Rn− (z) and Rn+ (z) given
by theorems 7.2.1 and 7.3.1 for the functions F − (z) and F + (z) respectively, that is,
n−1
1 X ±
±
F (z) = A (z)Φ± ± ±
k (ρ ; z) + Rn (z), (7.50)
m k=0 k
m
with ρ− = ρ = 1 − e−t0 and ρ+ = 1, and
k + λ±
± 1 ηz
Φk (c; z) := ± λ± Bc , ± m , k = 0, 1, 2, .... 0 ≤ c ≤ 1. (7.51)
(Λ ) m (Λ )
Therefore, the convergence character of expansion (7.47) comes from (7.27) and (7.38).
In other words, the dependence on z of the function h(t, z) does not spoil the convergence
character of the expansion. However, we must prove that it does not spoil its asymptotic
feature either.
First, we show that the terms between brackets inside the sum of (7.47) form an
k
±
asymptotic sequence, as z → ∞: From lemma 7.4.2 the coefficients Ak (z) = O z p , as
|z| → ∞. On the other side, from the asymptotic behavior of the beta and incomplete
beta functions [4, eqs. 5.12.1 and 5.11.12] and [116, eqs. 8.17.2, 8.17.4 and 8.18.3] we
find that the terms of the expansion
between brackets inside the sum in (7.47) are of
k k+λ+ k k+1
the order O z p − m + z p− m , as |z| → ∞. In other words, they form an asymptotic
sequence that is not of Poincaré type, but that decreases in the form of a sawtooth.
Secondly, we have to prove that the remainder Rn (z) satisfies the estimate given in the
theorem, as |z| → ∞. The key point to prove this is to split both integrals F ± (z) at the
point t = |z −1/p |, for |z| > x0 large enough. That is, we write F ± (z) = F0± (z) + F1± (z),
120 Chapter 7. A convergent Laplace’s method for integrals
with
Z |z −1/p |
−zηtm
F0± (z) = e h(t0 ± t, z)(t0 ± t)a−1 dt,
0
Z b±
(7.52)
± −zηtm a−1
F1 (z) = e h(t0 ± t, z)(t0 ± t) dt.
|z −1/p |
We remark that, for |z| large enough, we have that |z −1/p | ≤ t0 for any positive t0 .
Now, the integral F0− (z) is of the form of F − (z) with t0 replaced by |z −1/p |, and the
same applies to the integral F0+ (z) if we also replace h(t0 − t, z)(t0 − t)a−1 by h(t0 +
t, z)(t0 + t)a−1 . Therefore, we can apply theorems 7.2.1 or 7.3.1 to these two integrals
and we get
n−1
1 X ± −m/p |
F0± (z) = A (z)Φ± ±
k (ρz ; z) + Rn,0 (z), ρz := 1 − e−|z , (7.53)
m k=0 k
with Φ±
k (c; z) given in (7.51) and
Z (ρz )1/m ηz
± −1 λ± −1 ±
Rn,0 (z) := (1 − xm ) (Λ± )m x rn (x, z)dx, (7.54)
0
where rn± (x, z) is the n−th order Taylor remainder of h̃± (x, z) at x = 0.
Then, on the one hand, according to the splitting described above, we have
n−1
1 X ±
±
F (z) = F0± (z) + F1± (z) = Ak (z)Φ± ± ±
k (ρz ; z) + Rn,0 (z) + F1 (z). (7.55)
m k=0
On the other hand, we have formula (7.50). Then, from (7.50) and (7.55) we find that
Rn± (z) = Ψ± ± ±
n (z) + Rn,0 (z) + F1 (z), (7.56)
with
n−1
1 X ±
Ψ±
n (z) := A (z)[Φ± ± ±
k (ρ ; z) − Φk (ρz ; z)]. (7.57)
m k=0 k
In the remaining of the proof we study the asymptotic behavior of every one of the
± ±
three terms Ψ±n (z), Rn,0 (z) and F1 (z) on the right hand side of formula (7.56), in order
to find out the asymptotic behavior of Rn± (z), and then of Rn (z) = Rn+ (z) + Rn− (z).
Ψ±n (z). Note that the arguments of the two incomplete beta functions on the right
hand side of (7.57) (see (7.51)) are the same, the incomplete beta functions only
differ in their index. Then, taking into account the integral representation of the
incomplete beta function [4, eq. 8.17.1] we find that
Z ρ±
± ± ± 1 k+λ± ηz
−1
Φk (ρ ; z) − Φk (ρz ; z) = ± λ± t m −1 (1 − t) (Λ± )m dt.
(Λ ) ρz
Hence
Z 1
1
k+<λ± < ηz
−1
Φ± ±
k (ρ ; z) − Φ±
k (ρz ; z) ≤ t m
−1
(1 − t) (Λ± )m dt.
|(Λ )λ± |
±
1−e −|z −m/p |
7.4 A convergent and asymptotic Laplace’s method 121
−m/p |
Performing the change of variables t 7→ u defined by t = 1 − e−|z u we find
< ± m ηz
1
−|z −m/p |
Φ± ± ± (Λ )
k (ρ ; z) − Φ k (ρ z ; z) ≤ e ×
|(Λ± )λ± |
Z 1 −m
1 1
k+<λ± ηz ηz p |
−|z −m/p | −1 < (Λ± )m −1 −< ± )m |z
(1 − e u) m u du ≤ ±
e (Λ .
0 |(Λ± )λ | < ηz
(Λ± )m
(7.58)
Thus, from the asymptotic behavior of the coefficients A±k (z) (see lemma 7.4.2) and
formulas (7.57) and (7.58) we have that, for all k = 0, 1, . . . , n − 1,
−m
ηz
± bn/pc−1 −< (Λ± )m |z p |
Ψn (z) = O z e , as |z| → ∞. (7.59)
F1± (z). We have that f (t0 ± t) − f (t0 ) = ηtm + O(tp ) as t → 0+ with η > 0, and
therefore ∃ δ ± > 0 independent of t (and of course of z) such that
η
f (t0 ± t) − f (t0 ) − tm > 0 for 0 < t < δ ± . (7.60)
2
On the other hand, as t0 is an absolute minimum of f (t) in [0, ∞), there exist
± > 0 such that
f (t0 ± t) − f (t0 ) − ± > 0 for t ≥ δ ± . (7.61)
For large enough |z| we have that δ ± > |z −1/p | and we can split the integral F1± (z)
at t = δ ± . We write
F1± (z) = G± ±
1 (z) + G2 (z),
with Z δ±
m
G±
1 (z) := e−zηt e−zfm (t0 ±t) g(t0 ± t)(t0 ± t)a−1 dt,
|z −1/p |
Z b±
m
G±
2 (z) := e−zηt e−zfm (t0 ±t) g(t0 ± t)(t0 ± t)a−1 dt.
δ±
with K and K̄ positive constants independent of |z|. From the asymptotic behavior
of the incomplete gamma function [4, eq. 8.11.2] we deduce that
m−1 m
± −1 − η2 z 1− p
G1 (z) = O z p e , as |z| → ∞. (7.62)
122 Chapter 7. A convergent Laplace’s method for integrals
And then,
Z b±
e−<z[f (t0 ±t)−f (t0 )− ] |g(t0 ± t)(t0 ± t)a−1 |dt.
±
−± <z
|G±
2 (z)| ≤e
δ±
Using (7.61), we find that, for <z > x0 , e−<z[f (t0 ±t)−f (t0 )− ] ≤ e−x0 [f (t0 ±t)−f (t0 )− ] .
± ±
Taking also into account that the last integral above, with <z replaced by x0 , is
convergent by hypothesis H7.4.(vi), we conclude that
± <z
|G±
2 (z)| ≤ Ke
−
,
with K > 0 independent of |z|. From this formula and (7.62) we find
η 1− m ±
p
F1 (z) = O e− 2 z + e−z , as |z| → ∞. (7.63)
Rn,0
± ±
(z). Recall the integral representation of Rn,0 (z) given in (7.54). The factor
rn (x, z) is the Taylor remainder of h̃ (x, z) at x = 0. Then, rn± (x, z) admits the
± ±
λ± /m−1
log(1 − wm )
± −1
ϕ(w) := g(t0 ± t(w)) − (t0 ± t(w))µ .
wm
The functions ψ(w) and ϕ(w) are analytic in the disk D0 (r) that contains the circle
C. Then, due to the choice of the radius of the circle C to cancel out the variable z
in the exponent of the exponential of the function h̃± (w, z), we have that
−1/p )p | |ψ(w)| p |ψ(w)|
|h̃± (w, z)| ≤ e|z(2z |ϕ(w)| = e2 |ϕ(w)| ≤ K̄,
7.4 A convergent and asymptotic Laplace’s method 123
for some constant K̄ > 0 independent of |z|. For w ∈ C we also have |w−x| ≥ |z −1/p |
and |w|n = 2n |z −n/p |. Then
|rn± (x, z)| ≤ Kxn |z n/p |,
with K > 0 independent of x and |z|. Then, from (7.54) we have
1/m
Z ρz
ηz
± m < (Λ± )m −1 <λ± −1
|Rn,0 (z)| ≤ (1 − x ) x |rn± (x, z)|dx
0
1/m
Z ρz
ηz
n/p m < (Λ± )m −1 n+<λ± −1
≤ K|z | (1 − x ) x dx
0
n + <λ±
n/p ηz
= K|z |Bρz ,< .
m (Λ± )m
Using the asymptotic behavior of the incomplete beta function [4, eqs. 8.17.2,
8.17.4, 8.18.3] we find
±
± n
− n+λ
Rn,0 (z) = O z p m , as |z| → ∞. (7.64)
and then, from the relation Rn (z) = Rn− (z) + Rn+ (z), we get
+ −
n
− n+λ n
− n+λ
Rn (z) = O z p m +z p m + exp. small terms, as |z| → ∞,
with
1 1
z2
Z Z
z2 (t−1)2 z2 2 z2
a−1/2 −z 2 a−1/2 − z2 t2
U1 (a, z) := e 2 t e 2 dt = e 2 (1 − t) e dt = e Ū1 a,
2 ,
0 0 2
and
∞ ∞
z2
Z Z
z2 (t−1)2 z2 2 z2
a−1/2 −z 2 a−1/2 − z2 t2
U2 (a, z) := e 2 t e 2 dt = e 2 (t + 1) e dt = e Ū2
2 a, ,
1 0 2
with Ū1 (a, z) and Ū2 (a, z) defined in examples 3 and 4 respectively. Therefore, a con-
vergent and asymptotic expansion of the parabolic cylinder function U (a, z) for large
negative z follows from expansions (7.31) and (7.41).
Chapter 8
A Systematic “Saddle Point Near
An End Point” Asymptotic Method
The “saddle point near an end point” problem consists on finding an asymptotic
expansion of integrals of the form
Z b
F (x; α) = e−xf (t;α) g(t)(t − a)s−1 dt, α ∈ (α1 , α2 ), (8.1)
a
for large x uniformly valid in α. In this formula (a, b) and (α1 , α2 ) are finite or infinite
real intervals, α is a real parameter and f (t) and g(t) are smooth enough functions in
(a, b). We assume that a > −∞ and for the sake of generality, we let an integrable
singularity of the integrand at t = a given by the factor (t − a)s−1 , with s > 0.
For fixed α, the problem of finding an asymptotic expansion of F (x) for large x has
already been discussed in section 2.2.2 of chapter 2: According to Laplace’s method, for
large x, the main contribution of the integrand to the integral (8.1) should come from
the neighborhood of the points of the integration interval [a, b] where the phase function
f (t) attains its absolute minima. Without loss of generality we may assume that f (t)
has only one absolute minimum that occurs at a point t = t0 ∈ [a, b). Thus, in section
2.2.2 we have derived three different expansions (2.18), (2.19) or (2.21) depending on the
location of t0 with respect to the integration interval [a, b): (i), the absolute minimum
of f (t) occurs at the interior point t0 ∈ (a, b), with f 0 (t0 ) = 0 and f 00 (t0 ) > 0; (ii), the
absolute minimum of the phase function f (t) occurs at one of the end integration points,
say t = a, and it is a simple, critical point of f (t); or (iii), the function f (t) is strictly
incresing in [a, b) and therefore attains its absolute minimum at the end point t = a, with
f 0 (a) > 0.
125
126 Chapter 8. A systematic “saddle point near an end point” asymptotic method
The three situations are entirely different and lead to three formally different expan-
sions (2.18), (2.19) or (2.21). This does not generate any problem as far as the absolute
minimum t0 of f (t) is fixed. However, now the phase function f (t; α) is also a function
of a certain parameter α. Imagine that as α crosses a certain critical value α∗ ∈ (α1 , α2 ),
the absolute minimum of f (t) changes of nature from being an interior point to being
the end point t = a (or vice versa). We say that there is “a saddle point near an end
point”. In this case, which approximation (2.18), (2.19) or (2.21) should be used to esti-
mate the asymptotic behavior of F (x; α), for large x? Comparing the three expansions,
we can see that there is a formal discontinuous transition between the approximants of
F (x; α) given by (2.18), (2.19) and (2.21) even when the function F (x; α) is a continuous
function of the parameter α. But moreover, expansions (2.18) and (2.21) are useless
from a numerical point of view when α is close to α∗ because the coefficients hn of those
expansions blow up as α → (α∗ )+ and α → (α∗ )− respectively. We face the problem
of finding an asymptotic expansion valid for any value of the parameter α, that is, of
finding an expansion uniformly valid for α ∈ (α1 , α2 ). This type of approximations have
already been studied in the literature [13, 145, 154] and belong to the class of the so
called uniform asymptotic expansions.1 .
In the case of the integral (8.1) the origin of the discontinuous behavior of the asymp-
totic approximations (2.18), (2.19) and (2.21) is different: as the parameter α crosses the
critical value α∗ , the absolute minimum t = t0 (α) of the phase function crosses the end
integration point t = a and is no longer a critical point of f (t). Therefore, when t0 (α) is
inside the integration interval we use the change of variables f (t; α) − f (t0 (α); α) = u2 to
derive the asymptotic expansion (2.18) of F (x; α). On the other hand, when t0 (α) = a
we use another change of variables defined by f (t; α) − f (t0 (α); α) = u to deduce the ex-
pansion (2.21) of F (x; α). In order to obtain an expansion valid for either t0 (α) ∈ (a, b)
or t0 (α) = a we should use a change of variables valid for both situations. Then, we
consider a change of variables t 7→ u that depends on α and that encodes the tran-
sition from one case to the other one. That p is, we consider the change of variables
u2
f (t; α) − f (t0 (α); α) = 2 − cu, with c := ± 2f (a; α) − 2f (t0 (α); α), where the ± sign
is taken accordingly to α > α∗ or α < α∗ [154, Ch. 7, §3]. After some computations, the
1
The concept of uniform expansion here is different from the one introduced in chapter 3 of this thesis.
There, uniform means that the error is bounded by a constant independent of the uniform variable in
a large domain. Here, uniform means that the expansion is valid when another variable (or parameter)
runs in a certain domain.
127
following compound asymptotic expansion can be derived [145, Ch. 22], [154, Ch. 7, §3]
" √ X ∞ 0
√ X ∞
#
Us (c x) an U (c x) b n
F (x; α) ∼ e−xf (a;α) + s(s+1)/2 , x → ∞, (8.2)
xs/2 n=0 xn x n=0
x n
with Ua (z) := Γ(a) exp (z 2 /4) U (a − 1/2, −z), where U (a, z) is a parabolic cylinder func-
tion [144], and an , bn are certain coefficients that may be computed by inverting the
two-point Taylor series at the points t = a and t = t0 (α) of the function u(t) implicitly
2
defined by the change of variables f (t; α) − f (a; α) = u2 − cu. Expansion (8.2) was first
found by Bleistein who introduced an integration by parts method [13] that was later
applied in other scenarios, like for example in the derivation of a uniform asymptotic
expansion for integrals with two coalescing saddle points.
In the standard Laplace’s method, in either case t0 (α) ∈ (a, b) or t0 (α) = a, the
asymptotic sequence is given by elementary functions: they are nothing but inverse
powers of the asymptotic variable x. In contrast, in the uniform method “saddle point
near an end point” the asymptotic sequence is given in terms of a special function: the
parabolic cylinder function. This function encodes the abrupt transition between the two
standard Laplace’s expansions (2.18) and (2.21) as the parameter α crosses the critical
value α∗ , ensuring that expansion (8.2) is valid for any value of α ∈ (α1 , α2 ).
On the other hand, in either of these expansions, the standard Laplace’s expansions
or the uniform asymptotic expansion, the computation of the coefficients hn or (an , bn )
is not straightforward, as it is neccesary to invert the Taylor series of a function defined
implicitly by a certain change of variables. We can not find general analytical formulas for
these coefficients in traditional textbooks about asymptotics [105, 145, 154], but rather
an indication on how to compute the few first. Nevertheless, some more or less explicit
representations of these coefficients can be found in the literature: Perron’s method gives
them in terms of derivatives of an explicit function whereas Wojdylo [152, 153] found
a recurrence formula in terms of partial ordinary Bell polynomials whose complexity
increases with the number of terms considered. On the other hand, based on Cauchy-
type integral representations, in [56] the authors have given a stable and efficient manner
to approximate the coefficients that appear in uniform asymptotic expansions of integrals
and in particular, the coefficients that appear in the “saddle point near an end point”
uniform asymptotic method.
A simplification of the standard method of Laplace (in the non-uniform case) has been
introduced in [74] and summarized in subsection 2.2.3 of this thesis. On the one hand,
the computation of the coefficients in this modified method is simpler and systematic,
given by formula (2.24). On the other hand, the computation of the asymptotic sequence
is as simple as in the standard Laplace’s method. The main idea is to split the phase
function f (t) into its asymptotically dominant monomial f (m) (t0 )/m!(t − t0 )m and a
subdominant remainder fm (t). Moreover, in [68] the same simplifying idea has been
applied to integrals with two asymptotically relevant points: a saddle point and a pole.
By avoiding the change of variables inherent to the classical uniform method “saddle point
near a pole” [145, Ch. 21], [154, Ch. 7, §2], this uniform method has been simplified
providing a simpler uniform asymptotic expansion. On the one hand, the asymptotic
sequence given in [68] is, as in the classical method, written in terms of error functions.
On the other hand, as in the modification of the method of Laplace [subsection 2.2.3],
the coefficients can also be computed by means of an explicit, closed formula.
128 Chapter 8. A systematic “saddle point near an end point” asymptotic method
The aim of this chapter of the thesis is to continue this line of research and derive
a new uniform asymptotic method “saddle point near an end point” whose coefficients
can be computed by means of a closed, explicit formula. Again, the main idea is to
2
avoid the change of variable t 7→ u defined by the relation f (t; α) − f (a; α) = u2 − cu
that characterizes the classical uniform method. Instead, we split the phase function into
its asymptotically dominant part and a subdominant remainder. In the cases analyzed
so far, the standard Laplace’s method [74], the uniform method “saddle point near a
pole” [68] and also in the convergent Laplace’s method [Chapter 7], that asymptotically
dominant part was a monomial of the form f (m) (t0 )/m!(t − t0 )m . However, as we will see
below, the situation in the uniform asymptotic method “saddle point near an end point”
is different, as the asymptotically dominant part is not a monomial, but a polynomial of
the second degree with two non-constant terms. Despite this extra difficulty, we will see
below that the coefficients can be computed straightforwardly by means of a systematic,
explicit formula and, as in the classical method, the asymptotic sequence is again given
in terms of parabolic cylinder functions.
where we assume that (a, b) is a real interval that may be finite or infinite, but with
a finite end point, say a ∈ R. We also consider that α ∈ (α1 , α2 ), for a real interval
(α1 , α2 ). We assume that the functions f (t) and g(t) are real valued and, for the sake of
generality, we allow a branch point at the point t = a, the factor (t − a)s−1 with s > 0.
We also consider that the function f (t) possesses a unique absolute minimum t0 in [a, b).
That absolute minimum may be an interior point t0 ∈ (a, b) or an end point t0 = a. But
the location of that absolute minimum may change with the parameter α. Imagine that
t0 is an interior point that, eventually, as α crosses the critical value α∗ ∈ (α1 , α2 ), the
point t0 crosses the end point of the integration domain t = a and therefore the absolute
minimum of f (t) in [a, b) does no longer occur in an interior point but at the end point
t0 = a. We are interested in deriving an asymptotic expansion of F (x; α) for large |x|
(with arg x ∈ (−π/2, π/2)) uniformly valid for α ∈ (α1 , α2 ).
We consider the following hypotheses for the functions f (t) and g(t):
H8.1.(i) The functions f (t; α), f 0 (t; α), f 00 (t; α) and g(t) are continuous real functions
of their variables in their respective domains, where primes denote derivatives
with respect to the variable t.
H8.1.(ii) By subdividing the range of integration in (8.3) and/or reversing the order of
integration if necessary, we may assume, without loss of generality, that the
function f (t; α) has a unique minimum at the point t = t0 (α) in [a, b). We
consider the step function
(
0 if α ≤ α∗ ,
θ(α) := (8.4)
1 if α > α∗ .
8.1 Preparatory results 129
t 0 (α)
a
α
α 1 α ∗ α 2
Figure 8.1: Typical portrait of the minimum t0 (α) of the phase function f (t; α) as a function
of the parameter α. For α < α∗ the absolute minimum of f (t, α) in the integration interval is
located at the end point t = a. As α crosses the critical value α∗ , the absolute minimum of
f (t; α) moves to the interior of the integration interval (a, b) and coincides with its only relative
minimum t¯0 (α).
In this formula, t¯0 (α) denotes the only relative minimum point of f (t; α) in
[a, b) (when it exists) that we assume to be a simple critical point of f (t). That
is, f 0 (t¯0 (α); α) = 0 and f 00 (t¯0 (α); α) > 0. We have that t¯0 (α∗ ) = a and we
assume that t¯0 (α) > a for α > α∗ and that the function t¯0 (α) is a continuous
function of α. Therefore, t0 (α) is also a continuous function of the parameter
α (see figure 8.1).
H8.1.(iii) For α < α∗ the function f (t; α) attains its absolute minimum in [a, b) at the
point t0 (α) = a and f 0 (a; α) > 0. We assume that f 00 (t0 (α); α) = f 00 (a, α) > 0
also for α < α∗ and not only for α ≥ α∗ as specified in hypothesis H8.1.(ii).
H8.1.(iv) Both functions f (t; α) and g(t) have a Taylor expansion at the point t = t0 (α)
with common radius of convergence ρ > 0, and the coefficients f (n) (t0 (α); α)
are continuous functions of α for n = 0, 1, 2, . . ..
H8.1.(v) The integral (8.3) is absolutely convergent for sufficiently large |x|, say <x >
x0 , for some fixed x0 > 0.
In the classical uniform expansion (8.2) the coefficients (an , bn ) are, somehow, the
coefficients of the two-point Taylor expansion at the points u = 0 (the end point) and
s−1
(u − c)/f 0 t(u)
u = c (the saddle point) of the function h(u, α) := g t(u) t(u) − a
in the integrand of (8.3) after the convenient change of variables f (t; α) − f (t0 (α); α) =
u2
2
− cu. They can be computed by reverting the Taylor series of the function u(t) defined
by the change of variables at the points t = a and t = t¯0 (α). Thus, closed, explicit
formulae for those coefficients are not given in traditional text book on asymptotics
[105, 145, 154]. But, in [145, Remark 22.2] it is suggested that, may be, a standard
130 Chapter 8. A systematic “saddle point near an end point” asymptotic method
Taylor expansion at only one point, say u = 0 or u = c, will provide easier formulas
for the coefficients (an , bn ). Motivated by this remark, we re-consider the spliting of
the phase function f (t; α) into an asymptotically dominant Taylor polynomial and a
subdominant remainder (see subsection 2.2.3). But now, there are two asymptotically
relevant points: either t = a if α ≤ α∗ or t = t¯0 (α) if α ≥ α∗ . Then, we consider the
Taylor expansion of the phase function f (t; α) at the moving point t0 (α) defined in (8.5),
that is asymptotically relevant for any value of α. Before we derive our main result we
need some definitions.
Definition. (Asymptotically dominant Taylor polynomial). We consider the second-
order Taylor polynomial of the phase function f (t; α) at the moving point t = t0 (α)
defined in (8.5):
p(t; α) := f (t0 (α); α) + a1 (α) (t − t0 (α)) + a2 (α) (t − t0 (α))2 . (8.6)
From hypotheses H8.1.(ii) and H8.1.(iii) we have that a2 (α) := 12 f 00 (t0 (α); α) > 0,
for all α ∈ (α1 , α2 ). We also have that a1 (α) := f 0 (t0 (α); α) ≥ 0, with a1 (α) = 0, for
α ≥ α∗ .
For the sake of simplicity, in the remaining of the chapter we will omit the dependence
of a1 (α), a2 (α), t0 (α), t¯0 (α) and θ(α) on the parameter α and simply write a1 , a2 , t0 , t¯0
and θ.
Definition. (Asymptotically subdominant remainder). We also consider the remainder
of the asymptotically dominant Taylor polynomial of the phase function
fp (t; α) := f (t; α) − p (t; α) , (8.7)
and we denote by q (q ≥ 3) the order of the next non-vanishing derivative of the phase
function f (t; α) at t = t0 after the second derivative.
Thus, we split the phase function to re-write integral (8.3) in the form
Z b
F (x; α) := e−xp(t;α) h(t, x; α)(t − a)s−1 dt, h(t, x; α) := e−xfp (t;α) g(t). (8.8)
a
On the one hand, the phase function is now simpler as it is just a polynomial p(t; α) of
the second degree. On the other hand, the factor h(t, x; α) also contains the asymptotic
variable x. As we may guess from similar situations [68, 74], [Ch.7 of this thesis] and we
will see below, the dependence on x of the function h(t, x; α) does not spoil the asymptotic
character of the expansion that we are going to derive.
From the definition of the function h(t, x; α) (8.8) and hypothesis H8.1.(iv) we have
that the function h(t, x; α) has a Taylor expansion at t = t0 :
n−1
X
h(t, x; α) = hk (x; α)(t − t0 )k + rn (t, x; α), (8.9)
k=0
valid in a certain disk of center t0 and radius ρ > 0, Dρ (t0 ). In this formula, rn (t, x; α)
is the Taylor remainder of h(t, x; α) at t = t0 and hn (x; α) are the Taylor coefficients.
These coefficients can be computed in the form
n
X
hn (x; α) = Ck (x; α) Bn−k (α), (8.10)
k=0
8.1 Preparatory results 131
where Cn (x; α) and Bn (α) are, respectively, the Taylor coefficients at t = t0 of the
functions e−xfp (t;α) and g(t):
∞
X ∞
X
e−xfp (t;α) = Cn (x; α)(t − t0 )n , g(t) = Bn (α)(t − t0 )n , t ∈ Dρ (t0 ). (8.11)
n=0 n=0
Moreover, the coefficients hn (x; α) can be computed in terms of the Taylor coefficients
at t = t0 of the functions e−xf (t) and g(t). Denote by An (x; α) the Taylor coefficients of
the function e−xf (t;α) at t = t0 :
∞
X
−xf (t;α)
e = An (x; α)(t − t0 )n , t ∈ Dρ (t0 ) (8.12)
n=0
(ii) The coefficients hn (x; α) can be computed using formula (8.10) and the following
recursive relation for the coefficients Cn (x; α):
C0 (x; α) = 1,
C (x; α) = C (x; α) = . . . = C (x; α) = 0,
1 2 q−1
n−q (n−k)
(8.14)
x X Ck (x; α) f (t 0 )
Cn (x; α) = − n , n = q, q + 1, q + 2, . . . ,
(n − k − 1)!
k=0
(iii) The coefficients hn (x; α) are polynomials in x of degree bn/qc. Therefore, hn (x; α) =
bn/qc
O x as x → ∞.
Proof. (i) Formula (8.13) follows from the very definition of h(t, x; α) in (8.8), the
definition of the coefficients An (x; α) (8.12) and Bn (α) (8.11), the equality fp (t; α) =
f (t; α) − a1 (t − t0 ) − a2 (t − t0 )2 and the Taylor series of the exponential function:
∞
aj x(t−t0 )j
X (aj x)n
e = (t − t0 )jn , j = 1, 2, t ∈ R.
n=0
n!
132 Chapter 8. A systematic “saddle point near an end point” asymptotic method
(ii) To derive the recurrence relation (8.14) we observe that the function e−xfp (t;α) sat-
isfies the differential equation
d −xfp (t;α)
e = xe−xfp (t;α) [a1 + 2a2 (t − t0 ) − f 0 (t)] .
dt
Introducing in this equation the expansion of e−xfp (t;α) given in (8.11), the Taylor
expansion of f 0 (t) at t = t0 and equating coefficients of equal powers we find the
relation (8.14) by taking into account the definition of a1 , a2 and q.
(iii) The recurrence relation (8.14) shows that the coefficients Cn (x; α) are polynomials
in x of degree bx/qc. Therefore, Cn (x; α) = O xbn/qc as x → ∞. Then, (iii)
follows from (8.10).
Remark 8.1.2. The index q plays the same role as the index p did in the modified Laplace’s
method of subsection 2.2.3 with a slight difference: the index q may depend on the pa-
rameter α. But, for any α ∈ (α1 , α2 ) we have that q ≥ 3 and, as we will see below, this
possible dependence does not have any effect on the asymptotic analysis of F (x; α).
introduced in (8.3) with the notation introduced in the previous section. Assume also that
hypotheses H8.1.(i)-H8.1.(v) hold. Then, for n = 1, 2, 3, . . . ,
n−1
!
X
F (x; α) = e−xf (t0 ;α) hk (x; α)Φk (x; α, s) + Rn (x; α) , (8.16)
k=0
where the coefficients hk (x; α) are given in lemma 8.1.1 and can be explicitly computed in
terms of the Taylor coefficients of e−xf (t;α) and g(t) at t = t0 . The asymptotic sequence
Φk (x; α, s) can be obtained from the following recurrence relation
a1 +2a2 (a−t0 ) √
n h io
a21
Φ0 (x; α, s) := Γ(s) s exp x2 a1 (t0 − a) − a2 (t0 − a)2 + 4a U s − 1/2, √
2a2
x ,
(2a2 x)
2 2
The remainder Rn (x; α) in (8.16) satisfies, for any fixed α ∈ (α1 , α2 ), as |x| → ∞,
Proof. We split the phase function f (t; α) in the integral F (x; α) (8.15) in the form (8.8).
We find Z b
2
F (x, α) = e−xf (t0 ;α) e−a1 x(t−t0 )−a2 x(t−t0 ) h(t, x; α)(t − a)s−1 dt. (8.22)
a
We replace expansion (8.9) for the function h(t, x; α) at t = t0 into (8.22) and interchange
summation and integration to find
" n−1 #
X
F (x; α) = e−xf (t0 ;α) hk (x; α)Φ̃k (x; α, s) + Rn (x; α) , (8.23)
k=0
where the coefficients hk (x, α) are the Taylor coefficients of the function h(t, x; α) at
t = t0 that are given in lemma 8.1.1, and
Z b
2
Φ̃k (x; α, s) := e−a1 x(t−t0 )−a2 x(t−t0 ) (t − t0 )k (t − a)s−1 dt, k = 0, 1, 2, . . . , (8.24)
a
Z b
2
Rn (x; α) := e−a1 x(t−t0 )−a2 x(t−t0 ) (t − a)s−1 rn (t, x; α)dt, (8.25)
a
Thus, we define
Z ∞
2
Φk (x; α, s) := e−a1 x(t+a−t0 )−a2 x(t+a−t0 ) (t + a − t0 )k ts−1 dt, k = 0, 1, 2, . . . , (8.27)
0
and (8.16) follows from (8.23) except for exponentially small terms.
134 Chapter 8. A systematic “saddle point near an end point” asymptotic method
Now, the first line of formula (8.17) follows directly from (8.27) by using the integral
representation of the parabolic cylinder function [144, eq. 12.5.1]. The second formula
of the recurrence relation (8.17) comes from (8.27) by replacing k by k + 1 and splitting
the factor (t + a − t0 )k+1 = (t + a − t0 )k (t + (a − t0 )). Finally, the explicit formula (8.18)
follows by considering the binomial expansion of the term (t + a − t0 )k in (8.27).
The asymptotic behavior of Φk (x; α, s) follows either by an application of the classi-
cal Laplace’s method to the integral (8.27) or by using the asymptotic behavior of the
parabolic cylinder function for large argument [145, Ch. 11, §§11.2 and 11.3] in the for-
mula (8.18) and some computations, in the different three cases: (i) α < α∗ and then
t0 = a and a1 > 0, (ii) α = α∗ and then t0 = a and a1 = 0, and (iii) α > α∗ and then
t0 > a and a1 = 0. Using either method, we get Φk (x; α, s) = O x−σ(k) for x → ∞,
with σ(k) = (k + s) in case (i), σ(k) = (k + s)/2 in case (ii), and σ(k) = (k + 1)/2 in
case (iii). In addition, we know from lemma 8.1.1(iii) that the coefficients hk (x; α) are
polynomials in x of degree bk/qc. Therefore, the sequence hk (x; α)Φk (x; α, s) constitutes
an asymptotic sequence, as x → ∞, and its behavior is the specified in (8.19).
It remains to prove that the remainder Rn (x; α) satisfies the asymptotic property
(8.20). The proof of this fact is a bit long and tedious, and resembles the proof of the
asymptotic character of expansion (7.47) in the asymptotic and convergent Laplace’s
method developed in chapter 7.
F (x; α) = e−xf (t0 ;α) [F0 (x; α) + F1 (x; α) + F2 (x; α)] , (8.28)
with Z t0 −θ|x−1/q |
F1 (x; α) := ex[f (t0 ;α)−f (t;α)] g(t)(t − a)s−1 dt, (8.29)
a
Z b
F2 (x; α) := ex[f (t0 ;α)−f (t;α)] g(t)(t − a)s−1 dt, (8.30)
t0 +|x−1/q |
and
Z t0 +|x−1/q |
F0 (x; α) := ex[f (t0 ;α)−f (t;α)] g(t)(t − a)s−1 dt
t0 −θ|x−1/q |
(8.31)
Z t0 +|x−1/q |
)2
= e−a1 x(t−t0 )−a2 x(t−t0 h(t, x; α)(t − a)s−1 dt,
t0 −θ|x−1/q |
where θ = θ(α) is the step function defined in (8.4). We note that if α ≤ α∗ we have
θ = 0 and t0 = a. Therefore, if α ≤ α∗ , F1 (x; α) = 0.
We replace the expansion (8.9) of h(t, x; α) at t = t0 in the integral F0 (x; α) and
interchange summation and integration. We find
n−1
X
F0 (x; α) = hk (x; α)Φ0k (x; α) + Rn0 (x; α), (8.32)
k=0
and Z t0 +|x−1/q |
2
Rn0 (x; α) := e−a1 x(t−t0 )−a2 x(t−t0 ) (t − a)s−1 rn (t, x; α)dt. (8.34)
t0 −θ|x−1/q |
In the remaining of the proof we show that Rn0 (x; α) is of the order indicated in the
right hand side of (8.20) and that F1 (x; α), F2 (x; α) and Ψn (x; α, s) are exponentially
small compared with Rn0 (x; α), as |x| → ∞. Therefore, from (8.35) we will conclude the
asymptotic behavior (8.20) for Rn (x; α), as x → ∞.
F1 (x; α): We only study the case α ≥ α∗ as F1 (x; α) = 0 otherwise. In this case
a1 = 0 and
Then,
with C1 and C2 positive constants independent of |x|. In the last inequality above
we have used that g(t) is continuous in (a, t0 ), that its absolute value is bounded
and also that (t − a)s−1 is integrable. Therefore,
a 1− 2
−ε1 x − 22 x q
F1 (x; α) = O e +e , as |x| → ∞. (8.37)
136 Chapter 8. A systematic “saddle point near an end point” asymptotic method
F2 (x; α). A similar analysis (with a1 ≥ 0) shows that we can find two numbers
Λ2 > 0 and ε2 > 0 independent of x such that
a 1− 1 a 1− 2
−ε2 x − 21 x q − 22 x q
F2 (x; α) = O e +e , as |x| → ∞. (8.38)
with a2 > 0 and a1 > 0 for α > α∗ and a1 = 0 for α ≤ α∗ . Therefore, for each
k ∈ N we have that
1− 1 1− 2
0 −a1 x q −a2 x q
Φk (x; α) − Φk (x; α) = O e , as |x| → ∞,
and, as hk (x; α) are polynomials in x of degree bk/qc (see lemma 8.1.1) we conclude
that
1− 2
q
Ψn (x; α) = O e−a2 x , as |x| → ∞, (8.39)
Rn0 (x; α). If |x| is large enough, the integration interval I := (t0 − θ|x−1/q |, t0 +
|x−1/q |) in the definition (8.34) of Rn0 (x; α) is contained in the disk of convergence
of the Taylor series of h(t, x; α) at t = t0 . In that case, the remainder rn (t, x; α)
admits the Cauchy’s integral representation
(t − t0 )n
I
h(w, x; α)dw
rn (t, x; α) = n
,
2πi C (w − t0 ) (w − t)
where the integration path C is any simple closed contour around the point w = t0
that contains the point w = t0 and w = t, for all t ∈ I, oriented in the positive
direction. We choose C to be the circle of center t0 and radius 2|x−1/q |. Then, for
any w ∈ C, |w − t0 | = 2|x−1/q | and |w − t| ≥ |x−1/q |, for all t ∈ I. Moreover, we
have that
q ˜
h(w, x; α) = e[f (w;α)−p(w;α)] g(w) = e−x(w−t0 ) f (w;α) g(w),
for a certain function f˜(w; α) analytic in a disk of center t0 that contains the path
C. Thus, for all w ∈ C
q |f˜(w;α)| qM
|h(w, x; α)| ≤ e−|x||w−t0 | |g(w)| ≤ e−2 1
M2 ≤ M3 ,
With the aim to further analyze this remainder, we distinguish the three different
cases α < α∗ , α = α∗ and α > α∗ .
– Case 1: α < α∗ . We have θ = 0 and t0 = a. Then, from (8.41)
Z a+|x−1/q |
e−<x[a1 (t−a)+a2 (t−a) ] (t − a)n+s−1 dt
2
0 n/q
|Rn (x; α)| ≤ Kn |x |
a
Z |x−1/q |
e−<x[a1 t+a2 t ] tn+s−1 dt
2
n/q
= Kn |x |
Z0 ∞
e−<x[a1 t+a2 t ] tn+s−1 dt = Kn |xn/q |Φn (<x, s; α),
2
≤ Kn |xn/q |
0
which follows from the definition of Φn (x, s; α) in (8.27) in the particular case
t0 = a. Taking into account the asymptotic behavior of Φn (x, s; α) shown
above, we find
Rn0 (x; α) = O xn/q−(n+s) ,
as |x| → ∞. (8.42)
and
Z t0 +|x−1/q |
2
H2 (x; α) := e−a2 <x(t−t0 ) (t − t0 )n (t − a)s−1 dt
t0
Z |x−1/q |
2
= e−a2 <xt tn (t + t0 − a)s−1 dt,
0
Now, we have
in other words, as t0 > a, the factor (t0 − a ± t)s−1 does not have any influence
on the asymptotic analysis of the integrals H1 (x; α) and H2 (x; α). Then, a
direct application of Laplace method gives
Z |x−1/q | Z ∞
−a2 <xt2 n 2
Hk (x; α) ∼ (t0 − a) s−1
e t dt ≤ (t0 − a) s−1
e−a2 <xt tn dt
0 0
n+1
s−1
(t0 − a) Γ 2
= n+1 n+1
,
2a2 2
(<x) 2
Remark 8.2.3. The right hand side of (8.16) is not a genuine Poincaré type asymptotic
expansion of the function F (x; α), as the asymptotic sequence is not of the form x−n .
The asymptotic behavior of the terms in the expansion do not decrease monotonically
with the order n of the approximation, but in the form of a sawtooth.
8.3 Examples 139
8.3 Examples
In this section we consider the hypergeometric confluent functions M (c, d, x) and U (c, d, x)
[103] for large d and x, and of the same order, that is, we consider the functions
M (c, x/α + c + 1, x) and U (c, αx + c + 1, x) for large x and fixed α and c. In [67]
the authors have found three different non-uniform expansions for those functions and
large x, according to whether α > 1, α = 1 or α < 1, with explicit formulas for the
coefficients. On the other hand, in [139] the author has obtained uniform asymptotic
expansions for M (c, d, x) and U (c, d, x) for large d and x, in terms of parabolic cylinder
functions, by applying the classical method due to Bleistein, but this method does not
offer an explicit general formula for the coefficients of the expansion. Then, we are going
to apply theorem 8.2.1 to derive new uniform asymptotic expansions of the functions
M (c, x/α + c + 1, x) and U (c, αx + c + 1, x) for large x and fixed c, uniformly valid for
α ∈ (0, ∞), with explicit formulas for the coefficients.
for large |d| and |x| and of the same order. More precisely, we take d = αx + c + 1 with
fixed α ∈ (0, ∞), independent of x. Then,
Z ∞ Z ∞
1 −xt c−1 1
U (c, αx + c + 1, x) = e t αx
(1 + t) dt = e−x[t−α log(t+1)] tc−1 dt,
Γ(c) 0 Γ(c) 0
( (
0, if α ≤ 1, 1 − α, if α ≤ 1,
f (t0 ; α) = a1 =
α − 1 − α log α, if α ≥ 1, 0, if α ≥ 1,
( (
α
, if α ≤ 1, −2α, if α ≤ 1,
a2 = 2
1
f (3) (t0 ; α) = −2
2α
, if α ≥ 1, α2
, if α ≥ 1.
In particular, the third derivative of f (t, α) at t = t0 does not vanish for any value of
α ∈ (0, ∞) and therefore q = 3. Moreover, hypotheses H8.1.(i)-H8.1.(v) are satisfied and
140 Chapter 8. A systematic “saddle point near an end point” asymptotic method
Table 8.1: Absolute value of the relative errors in the approximation of U (c, α x + c + 1, x)
for c = 1.3, arg x = π/3 with increasing |x| and several values of α obtained by truncating the
right hand side of (8.46) after 4 terms.
with
n
(−1)n n X n (−αx)k
An (x; α) = x k
,
n! k=0
k [(1 + t0 )x]
which follows from lemma 8.1.1(i). On the other hand, they can also be computed
recursively in the form
h (x; α) = 1, h1 (x; α) = 0, h2 (x; α) = 0,
0
n−3
αx X (−1)n−k+1
hn (x; α) = hk (x; α) , for n ≥ 3.
(1 + t )n−k
n k=0
0
Table 8.2: Absolute value of the relative errors in the approximation of U (c, α x + c + 1, x) for
c = 1.7, arg x = −π/4 with increasing |x| and several values of α obtained by truncating the
right hand side of (8.46) after 7 terms.
Table 8.3: Absolute value of the relative errors in the approximation of U (c, α x + c + 1, x) for
c = 2.1, x real and positive and several values of α obtained by truncating the right hand side
of (8.46) after 11 terms.
x
for large |d| and |x| and of the same order. More precisely, we take d = α
+ c + 1, with
α ∈ (0, ∞) fixed, independent of x. Then
( (
1
0, if α ≤ 1, α
− 1, if α ≤ 1,
f (t0 ; α) = 1−α+log α a1 =
α
, if α ≥ 1, 0, if α ≥ 1,
( (
1 2
if α ≤ 1,
, if α ≤ 1,
,
a2 = 2α
f (3) (t0 ; α) = α
2α, if α ≥ 1, 2α2 , if α ≥ 1.
In particular, the third derivative of f (t, α) at t = t0 does not vanish for any value of
α ∈ (0, ∞) and therefore q = 3. Moreover, hypotheses H8.1.(i)-H8.1.(v) are satisfied and
142 Chapter 8. A systematic “saddle point near an end point” asymptotic method
Table 8.4: Absolute value of the relative errors in the approximation of M (c, x/α + c + 1, x)
for c = 1.8, arg x = π/6 with increasing |x| and several values of α obtained by truncating the
right hand side of (8.47) after 4 terms.
where (c)k denotes the Pochhammer’s symbol [4, §5.2(iii)]. Furthermore, the coefficients
hn (x; α) may be computed by means of
n bk/2c j k−2j
X X a a1
hn (x; α) = xk−j 2 An−k (x; α),
k=0 j=0
j! (k − 2j)!
with n
xn X n ( −x
α k
)
An (x; α) = ,
n! k=0 k [(t0 − 1)x]k
which follows from lemma 8.1.1(i). On the other hand, they can also be computed
recursively in the form
h (x; α) = 1, h1 (x; α) = 0, h2 (x; α) = 0,
0
n−3
−x X hk (x; α)
hn (x; α) = αn
, for n ≥ 3.
(1 − t )n−k
0
k=0
which follows from lemma 8.1.1(ii). Finally, tables 8.4, 8.5 and 8.6 contain some numerical
experiments about the accuracy of the expansion on the right hand side of (8.47) for
several values of the asymptotic variable x and the uniform parameter α.
8.3 Examples 143
Table 8.5: Absolute value of the relative errors in the approximation of M (c, x/α + c + 1, x)
for c = 1.6, arg x = −π/5 with increasing |x| and several values of α obtained by truncating
the right hand side of (8.47) after 7 terms.
Table 8.6: Absolute value of the relative errors in the approximation of M (c, x/α + c + 1, x)
for c = 2.1, x real and positive and several values of α obtained by truncating the right hand
side of (8.47) after 11 terms.
Chapter 9
Conclusions and Future Work
This thesis presents new ingredients in the analytical approximation of integral transforms
and, in particular, of special functions having an integral representation. In particular, we
seek new methods to derive approximations in the form of a series of integral transforms
satisfying the following three properties:
(a) The expansions are convergent.
(b) The expansions are given in terms of elementary functions whose coefficients can be
straightforwardly computed by means of a simple, closed and systematic formula.
(c) The expansions are valid in a large (possibly unbounded) region D of the complex
plane that contains small values of the selected variable or a selected point, say z = 0.
The main conclusions of the work are given in the next section 9.1. The chapter ends
with section 9.2 that exhibits some possible future work related with the research of this
thesis.
9.1 Conclusions
In CHAPTER 3, under some mild conditions on the functions h(t, z) and g(t), we have
designed Ra general method of constructing new expansions of the parametric integral
1
F (z) = 0 h(t, z)g(t)dt satisfying properties (a), (b) and (c) above. In contrast, the
classical Taylor or asymptotic expansions are only valid for, respectively, small or large
values of |z|. Therefore, the uniform approximation proposed in this chapter provides
smaller errors in any Lp (D) norm than the Taylor or asymptotic approximations when
the domain D is large enough1 .
The convergence of the uniform approximations can be either of exponential or power
type, depending on whether the end points t = 0 and t = 1 of the integration interval
are regular points of g(t) or not. More precisely, let Rn (z) denote the remainder of the
uniform approximation when it is truncated after n terms. We have that Rn (z) = O(a−n ),
for some a > 1 if the integration interval [0, 1] is completely contained in the region of
convergence Dr of the (multi-point) Taylor expansion of g(t). On the other hand, if
(0, 1) ⊂ Dr but [0, 1] ( Dr , that is, t = 0 and/or t = 1 are singular points of g(t), the
1
We only consider those values of p for which the Lp (D) norm is finite.
145
146 Chapter 9. Conclusions and future work
remainder satisfies Rn (z) = O(n−δ ), for some δ > 0 that is related with the behavior of
the function g(t)h(t, z) at the end points t = 0 and/or t = 1.
In the worst scenario, when one or both end points of the integration interval are
singular points of g(t) the speed of convergence of the uniform expansions is not impres-
sive. Still, they are a nice numerical tool for the evaluation of integral transforms F (z)
or special functions admitting such an integral representation. Indeed, in CHAPTER 4
we have applied the theory introduced in the previous chapter to many special functions
of the mathematical physics. These examples illustrate the applicability of the theory of
uniform approximations. In particular, we have derived new convergent expansions of the
following special functions: the Struve function Hν (z); the Bessel function of the first kind
Jν (z); the incomplete gamma functions γ(a, z) and Γ(a, z); the confluent hypergeometric
M (a, b, z) and U (a, b, z) functions; the exponential integral E1 (z); the symmetric ellip-
tic integrals RF (x, y, z) and RD (x, y, z); the Gauss hypergeometric function 2 F1 (a, b, c; z)
and the incomplete beta function Bz (a, b).
In the general theory of chapter 3 we have considered the possibility of expanding the
function g(t) at multiple points. This may be necessary from a theoretical point of view
to assure that the domain of convergence of the Taylor series contains the integration
interval. However, in all the derived uniform expansions of chapter 4 the use of one-
point Taylor series was enough to satisfy this requirement. Therefore, for simplicity, we
have taken only one base point. In this manner we have derived expansions given in
terms of elementary functions and that hold in unbounded regions of the complex plane
that contain small values of |z|. In the examples analyzed in chapter 4 such regions are
horizontal strips, half-planes, or a region S(θ) that depends on a certain angle θ ∈ [π/2, π)
that, in the limit case θ → π, becomes the cutted complex plane C \ (−∞, −1].
We have compared the uniform expansions with the classical expansions derived from
Taylor series or asymptotic expansions by means of numerical tables 4.1-4.10 and figures
4.1-4.7, 4.9-4.11. In general, the numerical results show that the use of Taylor power series
is recommended over the uniform approximation when we want to obtain an approxima-
tion valid only for small values of the variable. Similarly, asymptotic expansions should
be used if an approximation only for large values of the variables is needed. However,
there exists a certain intermediate region where the uniform approximation performs
better than both, Taylor series and asymptotic expansions. But, moreover, uniform ex-
pansions provide approximations that are globally more satisfactory. Therefore, uniform
approximations are recommended when we are working with the function F (z) in a large
region of the complex plane that contains both, small and large values of the variable |z|.
For example, in spectral problems of atomic models in quantum mechanics, the com-
putation of the spectrum requires the knowledge of the analytic behavior of the eigen-
functions for both, large and small values of the radius. In this problem, the use of
Taylor or asymptotic expansions is useless as their validity is local. In contrast, uniform
approximations are valid in a sufficiently large set and therefore they may be useful to
compute the spectrum easily.
More generally, consider a certain computation, may be an integral or a differential
equation, that involves the function F (z), and in which the variable z runs in a large
domain D. In general, the complicated analytical expression of F (z) makes that calcu-
lation impossible, but we may just replace F (z) by its uniform approximation given in
terms of elementary functions. This replacement is valid in a large region of the complex
plane and, in most cases, it will be valid for the whole range of values of the variable
9.1 Conclusions 147
where the Laplace’s method is usually used to obtain an asymptotic expansion for large
z. This method is very powerful but has three main drawbacks2 :
The expansion has a local validity as it is only valid for large values of |z|.
The asymptotic expansion is divergent and then it is required to find accurate error
bounds and to study the optimal truncation term of the series.
The coefficients of the expansion are defined by reverting a certain series implicitly
given by a change of variable.
We have solved these problems at once by designing a convergent and asymptotic Laplace’s
method for integrals whose coefficients can be computed by means of an explicit, sim-
ple formula, in terms of generalized Bernoulli polynomials and the Taylor coefficients of
e−zf (t) and g(t) at the asymptotically dominant point of f (t), as z → ∞. The price to
pay is that the asymptotic sequence is not straightforward to compute as in the classical
Laplace’s method, but given in terms of (incomplete) beta functions.
The conditions on the function f (t) and g(t) to apply the convergent and asymptotic
Laplace’s method are similar to the ones required for the application of the classical
Laplace method, except that we require for the functions to be analytic in a certain
region Sm (∞, 0). However, performing some tricks, this restriction can be relaxed and
require, basically, the analyticity of both functions in a neighborhood of the point t0
where the phase function f (t) attains its absolute minimum.
Then, the designed convergent and asymptotic Laplace’s method for integrals provides
asymptotic expansions satisfying properties (a) and (c) at the beginning of the chapter,
but we have lost property (b). More precisely, although the coefficients are given by an
explicit formula, the asymptotic sequence is given in terms of incomplete beta functions.
and we have sought an asymptotic expansion of F (z; α) for large z uniformly valid as
α varies. For fixed value of the parameter α, the method of Laplace provides three
different asymptotic expansions for this integral, depending on the location of the absolute
minimum of the phase function with respect to the integration interval [a, b]. However,
the three expressions are formally different. Imagine now that the parameter α varies in
a way that changes the nature of the absolute minimum of f (t), from being an interior
point t0 ∈ (a, b) to being one of the integration end points, say t0 = a. Then, neither
of the three expansions are useful when α is near that transition point, because the
coefficients blow up. Moreover, that transition is not continuous, but abrupt. Then, the
classical uniform “saddle point near an end point” asymptotic method has been designed
to solve this problem. At the cost of considering the parabolic cylinder function as
asymptotic sequence, an expansion valid uniformly in α can be found. However, the
coefficients of the expansion are computed by reverting a certain series implictly given by
a change of variables. For this reason, simple and explicit formulas for those coefficients
2
By no means these drawbacks discredites the method of Laplace for the approximation of integrals
when z is large.
9.2 Future work 149
are not given in traditional text books on asymptotics. In chapter 8 a modification of this
method is considered, obtaining an explicit and systematic formula for the coefficients of
the expansion. On the other hand, as in the classical method, the asymptotic sequence
is given by parabolic cylinder functions.
confluent U function; for h(t, z) = (− log t)z we get the gamma function or the Riemann’s
zeta function, etc.
Typically, using the classical Taylor series of h(t, z) at t = 0 (in the case when
h(t, z) = h(tz)) we find a power series of F (z) whereas asymptotic expansions of F (z)
are found from the asymptotic expansion of g(t) at the asymptotically relevant point of
h(t, z). We can instead consider the multi-point Taylor expansions of h(t, z) to derive new
expansions because the lemniscates of convergence of the multi-point Taylor expansion
avoid the singular points of h(t, z) (that, in general, depend on z) in a more efficient
manner. Then, it could be that the new expansions for F (z) hold in large regions of the
complex plane and at the same time are faster than the classical Taylor or asymptotic
expansions.
In CHAPTER 8 we have derived a systematic “saddle point near an end point method”.
This result is a continuation of the line of research initiated by my research group in
[68, 74, 75] to systemise classical asymptotic methods. In [68] the uniform asymptotic
“saddle point near a pole” method has been simplified, obtaining a systematic and explicit
formula for the coefficients of the expansion, obtaining the error function as asymptotic
sequence (as the classical method does). In chapter 8 we have systemised the uniform
asymptotic method “saddle point near an end point”. Continuing in this line, it remains
to find a systematic “two coalescing saddle points” uniform asymptotic method.
151
Appendix A
The Generalized Bernoulli Polynomials
(α)
The generalized Bernoulli polynomials Bn (x) can be defined by means of the generating
function [36, §24.16], [88, Ch.6 ], [140, Ch.1 §1.1]
α ∞ (α)
t xt
X Bn (x) n
e = t , |t| < 2π, (A.1)
et − 1 n=0
n!
where α and x are arbitrary complex numbers. For x = 0 we obtain the generalized
(α) (α)
Bernoulli numbers Bn (0) ≡ Bn , also called Nørlund polynomials [36, eq. 24.16.9],
that are polynomials of degree n of the complex variable α.
In this thesis, we use them to compute the Taylor coefficients at an appropriate point
of the composite function that results after a logarithmic change of variables of the form
(7.21).
Lemma A.0.1. For any m ∈ N, let φ(t) be an analytic function in a disk Dr (0) centered
at t = 0 with radius r > 0. For any λ ∈ C, consider the function
λ 1/m !
log(1 − xm ) log(1 − xm )
φ̃(x) := − φ x − .
xm xm
Then, the n−th Taylor coefficient of φ̃(x) at x = 0, that we denote φ̃n , is given in terms
of the derivatives of φ(t) at t = 0 by means of the formula
n
bmc n
(λ+1+ m )
X (−1)k Bk (1) dn−km h i
φ̃n = φ(t) . (A.2)
k=0
k! (n − km)! dtn−km t=0
Proof. It is clear that the function φ̃(x) is analytic in the open disk D0 (ρ), for some
0 < ρ < 1. From Cauchy’s formula for the n−th derivative of an analytic function, we
have
I m
λ m
1/m !
log(1 − x ) log(1 − x )
I
n! φ̃(x) n! dx
φ̃(n) (0) = n+1
dx = − m
φ x − m
,
2πi C x 2πi C x x xn+1
where C is the boundary of the disk of center 0 and radius ρ <1, oriented in the positive
h i1/m
log(1−xm )
sense and does not contain any singularity of the function φ x − xm . We
153
154 Appendix A. The Generalized Bernoulli Polynomials
h −tm i1/m
consider the change of variables x 7→ t given by (7.22), that is, x = t 1−etm ,
with t > 0 for x > 0. This transformation maps the disk D0 (ρ) to a certain region
Sm ((− log(1 − ρm ))1/m , 0), with Sm (q, 0) defined in (7.23). We find
λ+1+ mn
−tm
I
(n) n! −tm dt
φ̃ (0) = φ(t)e ,
2πi γ e−tm − 1 tn+1
where γ is the boundary of the region Sm ((− log(1 − ρm ))1/m , 0), that is a closed curve
encircling t = 0 in the positive direction. We may choose ρ small enough in order to
assure that Sm ((− log(1 − ρm ))1/m , 0) ⊂ Dr (0). Then, we apply Cauchy’s theorem again
to find " λ+1+ mn #
n m
d m −t
φ̃(n) (0) = n φ(t)e−t . (A.3)
dt e−tm − 1
t=0
Taking into account the generating function of the generalized Bernoulli polynomials
(A.1) and the k−th derivative at x = 0 of a composite function of the form Ψ(x) = ψ(xm ),
we have that
" λ+1+ mn # (
dk m 0 if k 6= 0 (mod m),
−t m −t
e = k! (λ+1+ n
)
dtk e−tm − 1 (−1)k/m (k/m)! Bk m
(1) if k = 0 (mod m).
t=0 m
(A.4)
Then, if we apply Leibniz’s formula for the n−th derivative of a product in (A.3) and
we use (A.4) we find the desired result (A.2).
The case λ = 0 and m = 1 is specially interesting. We have the following corollary:
Proof. We have
dn h i
n d
n h i
An = φ (− log t) = (−1) φ (− log(1 − x))
dtn t=1 dtn x=0
(n+1) k
(n)
In [88, pp. 129] we find the relation Bk (1) = 1 − n
Bk . Therefore
n
(−1)n−k dn−k h
X 1 k (n)
i
An = 1− Bk φ(t) ,
k=0
k! (n − k)! n dtn−k t=0
and formula (A.5) follows from rearranging the terms of the summation by means of the
change of index j = n − k.
Appendix B
A Larger Domain for a Uniformly
Convergent Expansion of the
Symmetric Elliptic Functions
When arg(−u/z) ∈ / [0, − arg z], u = x, y, we can invoke Cauchy’s theorem to rotate the
integration contour [0, ∞e−i arg z ) to the contour [0, ∞). Besides, arg(−u/z) = arg(u/z)−
πsign (arg(u/z)), and therefore, the condition arg(−u/z) ∈ / [0, − arg z] is equivalent to
the condition | arg(u/z) + arg z| < π. Taking into account that
(
| arg u| (< π) if | arg u − arg z| < π,
| arg(u/z) + arg z| =
| arg u ± 2π| (≥ π) if | arg u − arg z| ≥ π,
155
156 Appendix B. A larger domain for the uniform expansion of the SE functions
Im(x) Im(x)
z
Re(x) Re(x)
z
Im(x) Im(x)
Im(x)
Re(x) Re(x)
Figure B.1: The argument of the variable x is restricted to the sector arg x ∈ (arg z − π, arg z +
π) ∩ (−π, π]. The green region in all the pictures shows the different shapes of the x−section
of the region Λ1 for different arguments of z.
that is, the connection formula (4.20) is valid not only for z > 0, but in the bigger domain
Λ1 := {(x, y, z) ∈ (C \ (−∞, 0])3 : | arg x − arg z| < π, | arg y − arg z| < π}. (B.1)
In figure B.1 we illustrate the shape of the x−domain Λ1 for fixed z (the y−domain
for fixed z is analogous).
Similar arguments can be applied to the symmetric elliptic integral RD (x, y, z) to
find that the connection formulas (4.21) and (4.22) are respectively valid in the bigger
domains Λ1 and Λ2 , with Λ1 defined in (B.1) and
Λ2 := {(x, y, z) ∈ (C \ (−∞, 0])3 : | arg y − arg x| < π, | arg z 3 − arg x3 | < π}.
Therefore, uniform expansions for the symmetric elliptic integrals RF (x, y, z) and
RD (x, y, z) can be found from the connection formulas (4.20), (4.21) and (4.22) and the
expansion (4.37) not only when one of their variables is positive, but in a bigger domain.
In particular, the expansions derived for F (x, y) and G1 (x, y) hold for (x, y, z) ∈ Λ1
whereas the expansion derived for G2 (y, z) holds for (x, y, z) ∈ Λ2 .
Appendix C
The Integral Representation (6.4)
Of the p+1Fp(a, ~b, ~c; z) Function
In this appendix we describe the four different possibilities for the integration paths and
the constants A(bs , cs ) for all s = 1, 2, . . . , p in the integral representation (6.4) of the
generalized hypergeometric function p+1 Fp .
Γ(c)
A(b, c) = .
Γ(b)Γ(c − b)
Formula (6.4) is valid for <(ck ) > <(bk ) > 0 for all k = 1, 2, . . . , p.
(b) The integration contour L is a complex path that starts and terminates at t = 0. It
encircles the point t = 1 once in the positive direction. At the point where L cuts
the positive real axis we have arg(t) = 0 and arg(a − t) = π. This path is depicted
in figure C.1 (left). On the other hand,
(c) This case is, somehow, a mirror case of the previous one. In particular, the path L
starts and terminates at t = 1 and it encircles the point t = 0 once in the positive
direction. At the point where L cuts the negative real axis, we have that arg(t) = π
and arg(1 − t) = 0. In figure C.1 (right) this path is depicted. We also have
(d) The integration path starts and terminates at an arbitrary point P located on the
real axis between t = 0 and t = 1. It encircles the points 0 and 1 once in the positive
157
158 Appendix C. The integral representation (6.4) of the ~
p+1 Fp (a, b, ~
c; z) function
Figure C.1: The p identical integration contours L in (6.4), cases (b) and (c), may be deformed
to the respective contours (b) (left) and (c) (right) depicted in this figure. The horizontal
segments are assumed to be sticked to the real interval [0, 1] and the radius of the circles is
infinitesimally small. In case (b), at the point P we have that arg(t) = 0 and arg(1 − t) = π.
At the point P of figure (c) we have that arg(t) = π and arg(1 − t) = 0.
Figure C.2: The p identical integration contours L in (6.4), case (d), may be deformed to the
contour depicted in this figure. The horizontal segments are assumed to be sticked to the real
interval [0, 1] and the radius of the circles is infinitesimally small. At the point P we have
that arg(t) = arg(1 − t) = 0.
direction and then once in the negative direction. At the point P of L, we have
arg(t) = arg(1 − t) = 0. This path is depicted in figure C.2. On the other hand
eiπ(1−c) Γ(c)
A(b, c) = .
4Γ(b)Γ(c − b) sin(πb) sin[π(c − b)]
We assume that the contours L described in cases (b), (c) and (d) are squeezed
(analytically deformed) as much as possible around the real interval [0, 1], as it is shown
in figures C.1 and C.2.
The multiple integral representation (6.4) is a generalization of the respective integral
representation of the Gauss hypergeometric function 2 F1 (a, b, c; z) given in [104, eqs.
15.6.1, 15.6.2, 15.6.4 or 15.6.5]. The derivation of these multiple integral representation
is a straightforward generalization of the derivation of the corresponding simple integral
representation of the 2 F1 function. In particular, for any case (a)-(d) we replace the
159
expansion
∞
−a
X (a)n
(1 − T z) = (zT )n , |zT | < 1,
n=0
n!
into the right hand side of (6.4), interchange summation and integration and use the
reflection formula of the gamma function. We also use the different integral representa-
tions of the beta function [4, eqs. 5.12.1, 5.12.10 or 5.12.12] depending on the case (a),
(b), (c) or (d) under consideration:
1, case (a),
iπb
Z
Γ(a)Γ(b) 2ie sin(πb), case (b),
ta−1 (1 − t)b−1 dt = ×
L Γ(a + b)
2ieiπa sin(πa), case (c),
−4eiπ(a+b) sin(πa) sin(πb),
case (d).
Then, we obtain the series definition (6.2) of the generalized hypergeometric function
~ c; z). Therefore, the right hand side of (6.4) represents the analytic continua-
p+1 Fp (a, b, ~
tion in the variable z of p+1 Fp (a, ~b, ~c; z), defined by the right hand side of the series (6.2),
from the disk |z| < 1 to the cutted complex plane | arg(1 − z)| < π.
Appendix D
The Coefficients of the New Expansions
For the p+1Fp Function
In this appendix we give a recursive formula to compute several coefficients that appear
in the expansions (6.5), (6.18) and (6.25) of chapter 6 for the generalized hypergeometric
p+1 Fp function.
Lemma D.0.1. The functions p+1 Fp (−k, ~b, ~c; z) are polynomials in the variable z of degree
k. They can be computed using the following recurrence relation, valid for k = 1, 2, . . .
!
0, ~b
F z = 1,
p+1 p
~c
~
!
~
! p
! −−−→ !
−k, b 1 − k, b Y b s 1 − k, b +1
p+1 Fp z = p+1 Fp z −z p+1 Fp −−−→ z ,
~c ~c c
s=1 s c + 1
(D.1)
−−−→
where b + 1 := (b1 + 1, b2 + 2, . . . , bp + 1).
Proof. Using the series definition [3, 16.2.1] it is straightforward to check that the function
~ c; z) is a polynomial in the variable z of degree k as the defining series
p+1 Fp (−k, b, ~
terminates after k + 1 terms. On the other hand, the three term recurrence relation
follows from the integral representation (6.4) in the particular case when a = −k, with
k ∈ N, by splitting the factor (1 − zT )k = (1 − zT )k−1 (1 − zT ) and performing some easy
computations.
Lemma D.0.2. A recursive formula for the coefficients A2n (a, z) and Bn2 (a, z) in the two-
point Taylor expansion (6.20) of f (T ) = (1 − zT )−a at the points T = q and T = 1 − q
with q ∈ [0, 1/2) is the following:
161
162 Appendix D. The coefficients of the new expansions for the p+1 Fp function
and
2 2
(z, n, a, q)A2n (a, z) + M12 (z, n, a, q)Bn2 (a, z)
2 M11
An+1 (a, z) = ,
(n + 1)(4q 2 − 4q + 1) [z − 1q(1 − q)z 2 ]
(D.3)
M 2 (z, n, a, q)A2n (a, z) + M22
2
(z, n, a, q)Bn2 (a, z)
2
= 21
Bn+1
(a, z) ,
(n + 1)(4q 2 − 4q + 1) [z − 1q(1 − q)z 2 ]
where
2
(z, n, a, q) := z(a + 2n) 1 − (1 + 2q 2 − 2q)z ,
M11
2
M12 (z, n, a, q) := z 2 [−n + q(q − 1)(a + 1)] + z [2aq(1 − q) + 1 + 3n] − 2n − 1,
2
M21 (z, n, a, q) := −z(2 − z)(a + 2n),
2
(z, n, a, q) := z 2 n(1 + 4q − 4q 2 ) + 2(1 + a)q(1 − q) − (a + 2 + 6n)z + 4n + 2.
M22
(D.4)
Derivate this expression with respect to T to obtain the two-point Taylor expansion
of f 0 (T ) at the points T = q and T = 1 − q. Then, introducing both expressions
into the differential equation and equating the coefficients of [(T − q)(T + q − 1)]n and
T [(T − q)(T + q − 1)]n we obtain the recurrence scheme (D.2)-(D.3)-(D.4).
Lemma D.0.3. A recursive formula for the coefficients A3n (a, z), Bn3 (a, z) and Cn3 (a, z)
in the three-point Taylor expansion (6.28) of f (T ) = (1 − zT )−a at the points T = q,
T = 1/2 and T = 1 − q with q ∈ [0, 1/2) is the following:
3
M11 (z, n, a, q)A3n (a, z) + M12
3
(z, n, a, q)Bn3 (a, z) + M13
3
(z, n, a, q)Cn3 (a, z)
A3n+1 (a, z) = ,
2(n + 1)(1 − 2q)4 (z − 2) [(q − 1)qz 2 + z − 1]
M 3 (z, n, a, q)A3n (a, z) + M22
3
(z, n, a, q)Bn3 (a, z) + M23
3
(z, n, a, q)Cn3 (a, z)
3
Bn+1 (a, z) = 21 ,
2(n + 1)(1 − 2q)4 (z − 2) [(q − 1)qz 2 + z − 1]
M 3 (z, n, a, q)A3n (a, z) + M32
3
(z, n, a, q)Bn3 (a, z) + M33
3
(z, n, a, q)Cn3 (a, z)
3
Cn+1 (a, z) = 31 .
(n + 1)(1 − 2q)4 (z − 2) [(q − 1)qz 2 + z − 1]
(D.6)
163
3
(z, n, a, q) :=4 8q 4 z 2 (az + z − 2) + q 2 7(a + 1)z 3 + 6(2a − 1)z 2 − 12(a + 2)z + 16
M12
+q (a + 1)z 3 − 2(6a + 5)z 2 + 12(a + 2)z − 16 − 16q 3 z 2 (az + z − 2)
3
(z, n, a, q) :=n −12 + 28z − 21z 2 + 5z 3 − 4q 3z 3 + 42z 2 − 106z + 60 + 12q 2 19z 3
M13
+2z 2 − 46z + 20 + 16q 3 z 16 + 18z − 35z 2 + 8q 4 z 75z 2 − 18z − 16
3
(z, n, a, q) := 4z(a + 3n) 3 4q 2 − 4q − 1 z 2 + 8q 2 − 8q + 26 z − 24 ,
M21
3
36q 2 − 36q − 3 z 3 + 50 + 40q − 40q 2 z 2 − 96z + 48
M22 (z, n, a, q) :=6n
+ 8z a 4q 2 − 4q − 5 − 18 − 12z 2 a 4q 2 − 4q − 3 + 8q 2 − 8q − 4
3
8q 2 − 8q − 1 z 2 + 12z − 12 ,
M31 (z, n, a, q) := −4z(a + 3n)
3
(z, n, a, q) := − 6n 20q 2 − 20q − 1 z 3 − 24 q 2 − q − 1 z 2 − 48z + 24
M32
− 4 z 2 a −4q 2 + 4q + 5 − 12q 2 + 12q + 6 + 6(a + 1)(q − 1)qz 3
Proof. The proof of this lemma is identical to the proof of the previous one, but consid-
ering the three-point Taylor expansion of f (T ) at T = q, T = 1/2 and T = 1 − q:
∞
X
An (a, z) + Bn3 (a, z)T + Cn3 (a, z)T 2 [(T − q)(T − 1/2)(T + q − 1)]n .
3
f (T ) =
n=0
Bibliography
[2] A. Apelblat, Integral Transforms and Volterra Functions, Nova Science Publish-
ers, New York, 2010.
[4] R. A. Askey and R. Roy, Gamma Function, in: NIST Handbook of Mathe-
matical Functions, Cambridge University Press, Cambridge. 2010, pp. 135 – 147
(Chapter 5).
[6] A. Bejan, Convection Heat Transfer, Wiley, New York, 1984, doi:
10.1002/9781118671627.
[10] M. V. Berry and C. J. Howls, Hyperasymptotics, Proc. Roy. Soc. London Ser.
A, 430 (1990), pp. 653–668, doi: 10.1098/rspa.1990.0111.
[11] M. V. Berry and C. J. Howls, Axial and focal-plane diffraction catastrophe in-
tegrals, J. Phys. A: Math. Theor., 43 (2010), doi: 10.1088/1751-8113/43/37/375206.
[12] M. V. Berry and C. J. Howls, Integrals with Coalescing Saddles, in: NIST
Handbook of Mathematical Functions, Cambridge University Press, Cambridge.
2010, pp. 775 – 793 (Chapter 36).
165
166 Bibliography
[31] S. Colombo, Sur la fonction ν(t, n) (in french), C. R. Acad. Sci. Paris, 226 (1948),
pp. 1235–1236.
[32] S. Colombo, Sur les fontions ν(x, n) et µ(x, m, n), (in french), Bull. Sci. Math.,
2 (1955), pp. 72–78.
[33] P. J. Davis, Leonhard Euler’s Integral: A Historical Profile of the Gamma Func-
tion, Amer. Math. Monthly, 66 (1959), pp. 849–869, doi: 10.2307/2309786.
[35] F. F. Di Bruno, Théorie des Formes Binaires (in french), Librerie Breno, Torino,
1876.
[36] K. Dilcher, Bernoulli and Euler Polynomials, in: NIST Handbook of Mathe-
matical Functions, Cambridge University Press, Cambridge. 2010, pp. 587–599
(Chapter 24).
[42] C. Ferreira, J. L. López and E. Pérez Sinusı́a, The use of two-point Taylor
expansions in singular one-dimensional boundary value problems I, J. Math. Anal.
Appl., 463 (2018), pp. 708–725, doi: 10.1016/j.jmaa.2018.03.041.
[58] E. Landau, Vorlesungen über Zahlentheorie [3 volumes] (in german), Verlag von
S. Hirzel, Leipzig., 1927. English translation in 1996 by AMS Chelsea Publishing:
Elementary number theory.
[68] J. L. López and P. J. Pagola, A Systematic “Saddle Point Near a Pole” Asymp-
totic Method with Application to the Gauss Hypergeometric Funtion, Stud. Appl.
Math., 127 (2011), pp. 24 – 37, doi: 10.1111/j.1467-9590.2010.00510.x.
[78] J. L. López and E. Pérez Sinusı́a, Two-point Taylor expansions and one-
dimensional boundary value problems, Math. Comp., 79 (2010), pp. 2103–2115,
doi: 10.1090/S0025-5718-10-02370-7.
[79] J. L. López and E. Pérez Sinusı́a, New series expansions for the confluent
hypergeometric function M (a, b, z), Appl. Math. Comput., 235 (2014), pp. 26 – 31,
doi: 10.1016/j.amc.2014.02.099.
[83] J. L. López and N. M. Temme, New series expansions of the Gauss hypergeomet-
ric function, Adv. Comput. Math., 39 (2013), pp. 349–365, doi: 10.1007/s10444-
012-9283-y.
[84] P.-C. Lu, Introduction to the Mechanics of Viscous Fluids, Holt, Rinehart and
Winston, New York, 1973.
[90] K. E. Muller, Computing the confluent hypergeometric function M (a, b, x), Num-
ber. Math., 90 (2001), pp. 179 – 196, doi: 10.1007/s002110100285.
[91] G. Nemes, On the Coefficients of the Asymptotic Expansion of n!, J. Integer Seq.,
13 (2010), doi: 10.48550/arXiv.1003.2907.
[92] G. Nemes, An Explicit Formula for the Coefficients in Laplace’s Method, Constr.
Approx., 38 (2013), pp. 471–487, doi: 10.1007/s00365-013-9202-6.
[93] G. Nemes, Error bounds and exponential improvement for Hermite’s asymptotic
expansion for the gamma function, Appl. Anal. Discrete Math., 7 (2013), pp. 161–
179, doi: 10.2298/AADM130124002N.
[94] G. Nemes, Error bounds and exponential improvement for the asymptotic expan-
sion of the Barnes G−function, Proc. R. Soc. Lond. Ser. A. Math. Phys. Eng. Sci.,
470 (2014), doi: 10.1098/rspa.2014.0534.
[95] G. Nemes, The resurgence properties of the large-order asymptotics of the Han-
kel and Bessel functions, Anal. Appl. (Singap.), 12 (2014), pp. 403–462, doi:
10.1142/S021953051450033X.
[96] N. Nielsen, Recherches sur les séries de factorielles (in french), Ann. Sci. École
Norm. Sup., 3 (1902), pp. 409–453, doi: 10.24033/asens.515.
[99] N. E. Nørlund, Sur les séries de facultés (in french), Acta Math., 37 (1914),
pp. 327–387, doi: 10.1007/BF02401838.
[100] N. E. Nørlund, Hypergeometric functions, Acta Math., 94 (1955), pp. 289 – 349,
doi: 10.1007/BF02392494.
[102] A. B. Olde Daalhuis, On the asymptotics for late coefficients in uniform asymp-
totic expansions of integrals with coalescing saddles, Methods Appl. Anal., 7 (2000),
pp. 727–745, doi: 10.4310/MAA.2000.v7.n4.a7.
[116] R. B. Paris, Incomplete Gamma and Related Functions, in: NIST Handbook of
Mathematical Functions, Cambridge University Press, Cambridge. 2010, pp. 173 –
192 (Chapter 8).
[117] R. B. Paris, Struve and Related Functions, in: NIST Handbook of Mathematical
Functions, Cambridge University Press, Cambridge. 2010, pp. 287 – 301 (Chapter
11).
[122] M. Parodi, Sur les solutions d’équations intégrales dont les noyaux renferment des
polynomes d’Hermite (in french), C. R. Acad. Sci. Paris, 224 (1947), pp. 780–782.
R∞
[123] M. Parodi, Sur une propriété de l’équation intégrale 0 (tx f (x)/Γ(1+x))dx = g(t)
(in french), Ann. Soc. Sci. Bruxelles I, 62 (1948), pp. 24–26.
[125] O. Perron, Über die näherungweise Berechnung von Funktionen großer Zhalen (in
german), Sitzungsber. Bayr. Akad. Wissensch. (Münch. Ber., 11 (1917), pp. 191–
219.
[127] H. Poincaré, Sur les intégrales irrégulières: Des équations linéaires, Acta Math.,
8 (1886), pp. 295–344, doi: 10.1007/BF02417092.
[128] M. Razpet, An application of elliptic integrals, J. Math. Anal. Appl., 168 (1992),
pp. 425–429, doi: 10.1016/0022-247X(92)90170-I.
[129] B. Riemann, Theorie der durch die Gauss’sche Reihe F (α, β, γ, x) darstellbaren
Funktionen (in german), Abh. Königl. Ges. Wiss. Göttingen, 7 (1857), pp. 3–22.
[130] B. Riemann, Über die Anzahl der Primzahlen unter einer gegebenen Grösse (in
german), Ber. Akad. Wiss. Berlin, (1859).
[131] J. Riordan, Combinatorial identities, John Wiley & Sons, Inc., New York-
London-Sydney, 1968.
174 Bibliography
[145] N. M. Temme, Asymptotic Methods for Integrals, World Scientific, New Jersey,
2015, doi: 10.1142/9195.
[148] V. Volterra, Teoria delle potenze, dei logaritmi e delle funzioni di decompo-
sizione (in italian), Acc. Lincei., Memoire 5 (1916), pp. 167–249.
[150] G. N. Watson, The Harmonic Functions Associated with the Parabolic Cylinder,
Proc. London. Math. Soc., 2 (1918), pp. 241–246, doi: 10.1112/plms/s2-17.1.116.
[153] J. Wojdylo, On the coefficients that arise from Laplace’s method, J. Comput.
Appl. Math., 196 (2006), pp. 241–266, doi: 10.1016/j.cam.2005.09.004.
[154] R. Wong, Asymptotic Approximation of Integrals, Society for industrial and Ap-
plied Mathematics, Philadelphia, 2001, doi: 10.1137/1.9780898719260.
[155] M. Wyman and R. Wong, The Asymptotic behavior Of µ(z, β, α), Canadian J.
Math., 21 (1969), pp. 1013–1023, doi: 10.4153/CJM-1969-112-4.